feat(agent): strip raw tool-scaffolding from streamed answers — deterministic filter with one bounded recovery

2026-09-03 13:39:15 -04:00
parent 801639efcc
commit 575d6c88d0
38 changed files with 2793 additions and 50 deletions
@@ -0,0 +1,20 @@
 All criteria verified green; no defects found, no fixes needed.
 ## Phase 70 — Harness-Aligned Tools: verification report
 **Verified (final pass — all 5 tasks already complete in `complete/`):**
 - `AGENT_TOOLS` = exactly `ls`(path?) / `read`(path) / `grep`(pattern, path?); `rg "list_documents|read_document|search_documents" app/` → no matches
 - `read` splits combined `source/path` at first slash, full content (A7); `grep` locked A5 (fixed substring, case-insens., 20×200, locator-only); `ls` prints `source: X | path: Y | title: Z`
 - SSE `tool` frames: new names; `argument` = the single string passed (`grep`→pattern, else path) or null
 - Kill switch (`MAX_ROUNDS=0` → one `tools=None` request) and LOW/deflected byte-identity pinned by unit tests and green
 - `TOOLS_SECTION` rewritten for the new surface; frontend `read`/`grep`/`ls` render (old names still render for persisted chats); README §"Agent document tools (ls + read + grep)" updated
 **Gates (exact commands):**
 - `uv run pytest --cov=app` → exit 0, all pass; TOTAL coverage **99%** (>90%)
 - `uv run ruff check . && uv run pyright` → "All checks passed!" / "0 errors"
 - `uv run pytest tests/e2e/test_harness_aligned_tools.py -v --no-cov` → **3 passed** (ls→read, grep→read, wire-argument rule)
 - Isolated regressions: `test_agent_document_tools` **4 passed**, `test_agent_unlimited_tools` **4 passed**, `test_search_tool` **3 passed**, `test_chat_rag` **3 passed**
 - Commit `801639e` exists with the exact Commit-block message; phase dir in `.agent/phases/complete/`
 **Deviations:** none — no code changes required this pass.
 **Next pending phase:** `71_scaffolding_guardrails` (in `todo/`).
@@ -0,0 +1,75 @@
 ........................................................................ [  5%]
 ........................................................................ [ 10%]
 ........................................................................ [ 15%]
 ........................................................................ [ 20%]
 ........................................................................ [ 25%]
 ........................................................................ [ 31%]
 ........................................................................ [ 36%]
 ........................................................................ [ 41%]
 ........................................................................ [ 46%]
 ........................................................................ [ 51%]
 ........................................................................ [ 57%]
 ........................................................................ [ 62%]
 ........................................................................ [ 67%]
 ........................................................................ [ 72%]
 ........................................................................ [ 77%]
 ........................................................................ [ 83%]
 ........................................................................ [ 88%]
 ........................................................................ [ 93%]
 ........................................................................ [ 98%]
 ...................                                                      [100%]
 =============================== warnings summary ===============================
 .venv/lib/python3.13/site-packages/fastapi/testclient.py:1
  /var/home/ducoterra/Projects/Personal/brain_of_reese/.venv/lib/python3.13/site-packages/fastapi/testclient.py:1: StarletteDeprecationWarning: Using `httpx` with `starlette.testclient` is deprecated; install `httpx2` instead.
    from starlette.testclient import TestClient as TestClient  # noqa
 -- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
 ================================ tests coverage ================================
 _______________ coverage: platform linux, python 3.13.13-final-0 _______________
 Name                        Stmts   Miss  Cover
 -----------------------------------------------
 app/__init__.py                 1      0   100%
 app/api/__init__.py             0      0   100%
 app/api/auth.py                22      0   100%
 app/api/chat.py               149      0   100%
 app/api/chats.py              110      0   100%
 app/api/config.py               7      0   100%
 app/api/doc_drafts.py          93      0   100%
 app/api/docs.py                50      0   100%
 app/api/git_sources.py        212      0   100%
 app/api/health.py              10      0   100%
 app/api/steering.py            42      0   100%
 app/api/suggestions.py          8      0   100%
 app/api/sync.py               101      0   100%
 app/config.py                 126      0   100%
 app/core/__init__.py            0      0   100%
 app/core/auth.py               20      0   100%
 app/core/caching.py           108      0   100%
 app/core/debugging.py          29      2    93%
 app/core/docs_push.py          39      0   100%
 app/core/logging.py            13      0   100%
 app/db.py                      21      0   100%
 app/main.py                    52      0   100%
 app/models.py                  86      0   100%
 app/rag/__init__.py             0      0   100%
 app/rag/agent.py              139      0   100%
 app/rag/archive_upload.py     128      0   100%
 app/rag/chunker.py            206      4    98%
 app/rag/git_sources.py         14      0   100%
 app/rag/importer.py           180      3    98%
 app/rag/llm.py                213      0   100%
 app/rag/overview.py            71      0   100%
 app/rag/prompts.py             65      0   100%
 app/rag/retriever.py           94      3    97%
 app/rag/source_removal.py      41      0   100%
 app/rag/sources_meta.py        16      0   100%
 app/rag/suggestions.py         27      0   100%
 app/rag/summarizer.py          24      0   100%
 app/schemas.py                205      0   100%
 -----------------------------------------------
 TOTAL                        2722     12    99%
 coverage gate: app/ 99% (>90%) OK
 All checks passed!
 0 errors, 0 warnings, 0 informations
 validation OK
@@ -0,0 +1,20 @@
 Done — the empty `.err` is a harness runtime artifact (tracked-report convention, same as prior phases); working tree is clean.
 **Phase 70, task 05 complete**
 - `tests/e2e/mock_llm.py`: deterministic flows now emit `ls` (no args, `call_0`), `read` with the **combined** `source/path` (mock joins the two catalog fields — only parse change), `grep` with `pattern`; docstrings + phase-70 notes updated
 - Existing E2E suites updated to new frames: `test_agent_document_tools.py` (`ls`→`read`), `test_agent_unlimited_tools.py` (multi-read `ls`+2×`read`), `test_search_tool.py` (`grep` + pattern argument)
 - NEW `tests/e2e/test_harness_aligned_tools.py` (3 tests): READ flow (ls line → 📄 Reading line → quoted answer → done.sources incl. read doc → no raw `<|…|>`/`tool_call` markup in DOM), SEARCH flow (🔎 Searching for sentinel → answer), wire argument rule across both flows (ls→null, read→combined path, grep→pattern; names ⊆ {ls, read, grep})
 - README agent-tools section rewritten for `ls`/`read(path)`/`grep(pattern, path?)` (combined identity + A5 locator semantics); `.env.example` header comment refreshed
 - Old-name cleanup so the rg criterion holds: `app/api/docs.py` route fn → `list_indexed_documents`, `app/rag/llm.py` docstrings, `test_llm_client.py`, `test_llm_stream_teardown.py`, `test_sse_events.py`, `test_mock_tool_flow.py`, `test_api.py` docstring. Remaining matches: only the explicitly-marked legacy frontend branches/pins + the marked legacy-chat persistence pin + the mandated "old names absent" regression pins from tasks 01–02
 **Gates (all green):**
 - `uv run pytest` → exit 0; `uv run pytest --cov=app --cov-report=term-missing` → TOTAL **99%** (>90%)
 - `uv run ruff check .` → clean; `uv run pyright` → 0 errors
 - E2E in isolation: `test_harness_aligned_tools` 3 passed, `test_agent_document_tools` 4, `test_agent_unlimited_tools` 4, `test_search_tool` 3, `test_chat_rag` 3
 **Decisions:** mock `_tool_flow` tuple kept `(source, path)`; the join happens at the wire-emission point (keeps `test_mock_tool_flow.py` pins valid). `.env.example` included in the commit for atomicity (not in the block's path list, but task 05 touched it). Phase-dir move left to the harness per executor rules.
 **Commit:** `801639e feat(agent): align the document tools with the harness-trained shape — ls, read(path), grep(pattern, path?)` (55 files, `--no-gpg-sign`)
 **Next pending:** `71_scaffolding_guardrails/01_filter.md`
@@ -0,0 +1,75 @@
 ........................................................................ [  5%]
 ........................................................................ [ 10%]
 ........................................................................ [ 15%]
 ........................................................................ [ 20%]
 ........................................................................ [ 25%]
 ........................................................................ [ 31%]
 ........................................................................ [ 36%]
 ........................................................................ [ 41%]
 ........................................................................ [ 46%]
 ........................................................................ [ 51%]
 ........................................................................ [ 57%]
 ........................................................................ [ 62%]
 ........................................................................ [ 67%]
 ........................................................................ [ 72%]
 ........................................................................ [ 77%]
 ........................................................................ [ 83%]
 ........................................................................ [ 88%]
 ........................................................................ [ 93%]
 ........................................................................ [ 98%]
 ...................                                                      [100%]
 =============================== warnings summary ===============================
 .venv/lib/python3.13/site-packages/fastapi/testclient.py:1
  /var/home/ducoterra/Projects/Personal/brain_of_reese/.venv/lib/python3.13/site-packages/fastapi/testclient.py:1: StarletteDeprecationWarning: Using `httpx` with `starlette.testclient` is deprecated; install `httpx2` instead.
    from starlette.testclient import TestClient as TestClient  # noqa
 -- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
 ================================ tests coverage ================================
 _______________ coverage: platform linux, python 3.13.13-final-0 _______________
 Name                        Stmts   Miss  Cover
 -----------------------------------------------
 app/__init__.py                 1      0   100%
 app/api/__init__.py             0      0   100%
 app/api/auth.py                22      0   100%
 app/api/chat.py               149      0   100%
 app/api/chats.py              110      0   100%
 app/api/config.py               7      0   100%
 app/api/doc_drafts.py          93      0   100%
 app/api/docs.py                50      0   100%
 app/api/git_sources.py        212      0   100%
 app/api/health.py              10      0   100%
 app/api/steering.py            42      0   100%
 app/api/suggestions.py          8      0   100%
 app/api/sync.py               101      0   100%
 app/config.py                 126      0   100%
 app/core/__init__.py            0      0   100%
 app/core/auth.py               20      0   100%
 app/core/caching.py           108      0   100%
 app/core/debugging.py          29      2    93%
 app/core/docs_push.py          39      0   100%
 app/core/logging.py            13      0   100%
 app/db.py                      21      0   100%
 app/main.py                    52      0   100%
 app/models.py                  86      0   100%
 app/rag/__init__.py             0      0   100%
 app/rag/agent.py              139      0   100%
 app/rag/archive_upload.py     128      0   100%
 app/rag/chunker.py            206      4    98%
 app/rag/git_sources.py         14      0   100%
 app/rag/importer.py           180      3    98%
 app/rag/llm.py                213      0   100%
 app/rag/overview.py            71      0   100%
 app/rag/prompts.py             65      0   100%
 app/rag/retriever.py           94      3    97%
 app/rag/source_removal.py      41      0   100%
 app/rag/sources_meta.py        16      0   100%
 app/rag/suggestions.py         27      0   100%
 app/rag/summarizer.py          24      0   100%
 app/schemas.py                205      0   100%
 -----------------------------------------------
 TOTAL                        2722     12    99%
 coverage gate: app/ 99% (>90%) OK
 All checks passed!
 0 errors, 0 warnings, 0 informations
 validation OK
@@ -0,0 +1,15 @@
 Task complete — all criteria green.
 **Implemented**
 - `app/rag/scaffolding.py`: pure module (`import re` only) with `SCAFFOLD_PATTERNS` (observed span + 2 standalone siblings, non-greedy span) and `ScaffoldingFilter` — streaming state machine: `feed()` strips leftmost-complete matches repeatedly, holds the live tail (open span or longest proper-prefix suffix of the 3 opening literals), `flush()` emits the tail as-is (pinned EOF choice), `stripped_chars` read-only property.
 - `tests/unit/test_scaffolding_filter.py`: 61-test matrix — span at every start-token split offset (0..19) + mid-span/end-token splits, multi-span, standalone tokens (alone/embedded/char-by-char), look-alikes verbatim, partial markers at EOF, exact surrounding-text preservation.
 **Results**
 - `uv run pytest tests/unit/test_scaffolding_filter.py -v --no-cov` → 61 passed
 - `rg "import (os|sys|logging|app)" app/rag/scaffolding.py` → no matches
 - `uv run pytest --cov=app` → 1448 passed; `app/rag/scaffolding.py` 100%, TOTAL 99% (>90%)
 - `uv run ruff check .` → clean; `uv run pyright` → 0 errors
 **Decisions** — open-span defined as "start with no end token *after* it" (prevents scaffolding leak when a lone end token precedes a start); repeated leftmost-strip loop pins spec behavior where removing one match can expose another. No deviations.
 **Next pending task:** `02_llm_integration.md` (wire filter into `chat_stream`/`chat_stream_retried`).
@@ -0,0 +1,77 @@
 ........................................................................ [  4%]
 ........................................................................ [  9%]
 ........................................................................ [ 14%]
 ........................................................................ [ 19%]
 ........................................................................ [ 24%]
 ........................................................................ [ 29%]
 ........................................................................ [ 34%]
 ........................................................................ [ 39%]
 ........................................................................ [ 44%]
 ........................................................................ [ 49%]
 ........................................................................ [ 54%]
 ........................................................................ [ 59%]
 ........................................................................ [ 64%]
 ........................................................................ [ 69%]
 ........................................................................ [ 74%]
 ........................................................................ [ 79%]
 ........................................................................ [ 84%]
 ........................................................................ [ 89%]
 ........................................................................ [ 94%]
 ........................................................................ [ 99%]
 ........                                                                 [100%]
 =============================== warnings summary ===============================
 .venv/lib/python3.13/site-packages/fastapi/testclient.py:1
  /var/home/ducoterra/Projects/Personal/brain_of_reese/.venv/lib/python3.13/site-packages/fastapi/testclient.py:1: StarletteDeprecationWarning: Using `httpx` with `starlette.testclient` is deprecated; install `httpx2` instead.
    from starlette.testclient import TestClient as TestClient  # noqa
 -- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
 ================================ tests coverage ================================
 _______________ coverage: platform linux, python 3.13.13-final-0 _______________
 Name                        Stmts   Miss  Cover
 -----------------------------------------------
 app/__init__.py                 1      0   100%
 app/api/__init__.py             0      0   100%
 app/api/auth.py                22      0   100%
 app/api/chat.py               149      0   100%
 app/api/chats.py              110      0   100%
 app/api/config.py               7      0   100%
 app/api/doc_drafts.py          93      0   100%
 app/api/docs.py                50      0   100%
 app/api/git_sources.py        212      0   100%
 app/api/health.py              10      0   100%
 app/api/steering.py            42      0   100%
 app/api/suggestions.py          8      0   100%
 app/api/sync.py               101      0   100%
 app/config.py                 126      0   100%
 app/core/__init__.py            0      0   100%
 app/core/auth.py               20      0   100%
 app/core/caching.py           108      0   100%
 app/core/debugging.py          29      2    93%
 app/core/docs_push.py          39      0   100%
 app/core/logging.py            13      0   100%
 app/db.py                      21      0   100%
 app/main.py                    52      0   100%
 app/models.py                  86      0   100%
 app/rag/__init__.py             0      0   100%
 app/rag/agent.py              139      0   100%
 app/rag/archive_upload.py     128      0   100%
 app/rag/chunker.py            206      4    98%
 app/rag/git_sources.py         14      0   100%
 app/rag/importer.py           180      3    98%
 app/rag/llm.py                213      0   100%
 app/rag/overview.py            71      0   100%
 app/rag/prompts.py             65      0   100%
 app/rag/retriever.py           94      3    97%
 app/rag/scaffolding.py         50      0   100%
 app/rag/source_removal.py      41      0   100%
 app/rag/sources_meta.py        16      0   100%
 app/rag/suggestions.py         27      0   100%
 app/rag/summarizer.py          24      0   100%
 app/schemas.py                205      0   100%
 -----------------------------------------------
 TOTAL                        2772     12    99%
 coverage gate: app/ 99% (>90%) OK
 All checks passed!
 0 errors, 0 warnings, 0 informations
 validation OK
@@ -0,0 +1,17 @@
 Task 02 complete. All gates green.
 **Implemented**
 - `app/rag/llm.py`: `chat_stream(..., scaffolding: ScaffoldingFilter | None = None)` — content deltas fed through the caller-owned filter (empty clean result yields nothing), thinking never filtered, `None` keeps the byte-identical raw path; `flush()` tail is yielded as a content piece **before** tool-call materialization; phase-48 teardown untouched. `chat_stream_retried(..., scaffolding=None)` forwards the same filter to every attempt (safe by construction — a restart only happens pre-piece, i.e. filter never fed).
 - Tool materialization unified to a single post-loop emission point (the old in-loop `finish_reason="tool_calls"` early emission was removed) so the flushed tail always precedes `ToolCallPiece`s; identical wire behavior for all real/fixture stream shapes.
 - Tests: 7 new `chat_stream` filter tests + 2 retry pass-through tests in `tests/unit/test_llm_client.py` (span mid-stream, split across chunks, pure-scaffolding → zero pieces, thinking raw, `None` raw pin, flush-before-tools order, abandon-with-filter teardown, filter forwarded to every attempt, unfed-filter reuse).
 - Added the `scaffolding` kwarg to the five other `chat_stream` fakes (test_agent, test_chat_gate, test_agent_tools, test_chat_api) — required interface maintenance for the retry primitive to forward it.
 **Results**
 - `uv run pytest tests/unit/test_llm_client.py tests/unit/test_llm_stream_teardown.py -v --no-cov` → 60 passed
 - `uv run pytest --cov=app --cov-report=term-missing` → 1457 passed; `app/rag/llm.py` 100%, TOTAL 99% (>90%)
 - `uv run ruff check . && uv run pyright` → clean
 - E2E regressions in isolation: `test_harness_aligned_tools` (3), `test_agent_document_tools` (4), `test_chat_rag` (3) → all passed
 **Decisions:** single post-loop tool-materialization point (simpler; pins the content-before-tools order at both former emission points); no commit — the phase commit belongs to task 05.
 **Next pending task:** `.agent/phases/todo/71_scaffolding_guardrails/03_recovery_policy.md`
@@ -0,0 +1,77 @@
 ........................................................................ [  4%]
 ........................................................................ [  9%]
 ........................................................................ [ 14%]
 ........................................................................ [ 19%]
 ........................................................................ [ 24%]
 ........................................................................ [ 29%]
 ........................................................................ [ 34%]
 ........................................................................ [ 39%]
 ........................................................................ [ 44%]
 ........................................................................ [ 49%]
 ........................................................................ [ 54%]
 ........................................................................ [ 59%]
 ........................................................................ [ 64%]
 ........................................................................ [ 69%]
 ........................................................................ [ 74%]
 ........................................................................ [ 79%]
 ........................................................................ [ 84%]
 ........................................................................ [ 88%]
 ........................................................................ [ 93%]
 ........................................................................ [ 98%]
 .................                                                        [100%]
 =============================== warnings summary ===============================
 .venv/lib/python3.13/site-packages/fastapi/testclient.py:1
  /var/home/ducoterra/Projects/Personal/brain_of_reese/.venv/lib/python3.13/site-packages/fastapi/testclient.py:1: StarletteDeprecationWarning: Using `httpx` with `starlette.testclient` is deprecated; install `httpx2` instead.
    from starlette.testclient import TestClient as TestClient  # noqa
 -- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
 ================================ tests coverage ================================
 _______________ coverage: platform linux, python 3.13.13-final-0 _______________
 Name                        Stmts   Miss  Cover
 -----------------------------------------------
 app/__init__.py                 1      0   100%
 app/api/__init__.py             0      0   100%
 app/api/auth.py                22      0   100%
 app/api/chat.py               149      0   100%
 app/api/chats.py              110      0   100%
 app/api/config.py               7      0   100%
 app/api/doc_drafts.py          93      0   100%
 app/api/docs.py                50      0   100%
 app/api/git_sources.py        212      0   100%
 app/api/health.py              10      0   100%
 app/api/steering.py            42      0   100%
 app/api/suggestions.py          8      0   100%
 app/api/sync.py               101      0   100%
 app/config.py                 126      0   100%
 app/core/__init__.py            0      0   100%
 app/core/auth.py               20      0   100%
 app/core/caching.py           108      0   100%
 app/core/debugging.py          29      2    93%
 app/core/docs_push.py          39      0   100%
 app/core/logging.py            13      0   100%
 app/db.py                      21      0   100%
 app/main.py                    52      0   100%
 app/models.py                  86      0   100%
 app/rag/__init__.py             0      0   100%
 app/rag/agent.py              139      0   100%
 app/rag/archive_upload.py     128      0   100%
 app/rag/chunker.py            206      4    98%
 app/rag/git_sources.py         14      0   100%
 app/rag/importer.py           180      3    98%
 app/rag/llm.py                216      0   100%
 app/rag/overview.py            71      0   100%
 app/rag/prompts.py             65      0   100%
 app/rag/retriever.py           94      3    97%
 app/rag/scaffolding.py         50      0   100%
 app/rag/source_removal.py      41      0   100%
 app/rag/sources_meta.py        16      0   100%
 app/rag/suggestions.py         27      0   100%
 app/rag/summarizer.py          24      0   100%
 app/schemas.py                205      0   100%
 -----------------------------------------------
 TOTAL                        2775     12    99%
 coverage gate: app/ 99% (>90%) OK
 All checks passed!
 0 errors, 0 warnings, 0 informations
 validation OK
@@ -0,0 +1,18 @@
 All green. Task 03 is complete.
 **Implemented**
 - `app/rag/agent.py`: `CORRECTION_INSTRUCTION` (verbatim constant), `MalformedReplyError(LLMError)`, `AgentHolder.scaffold_stripped`; `run_agent` now runs every request (rounds, forced final, recovery) through a fresh `ScaffoldingFilter`, logs one capture warning per stripped span (200-char truncation), and on a scaffolding-only no-calls round runs exactly one `tools=None` recovery with the correction folded into the single system message; a second empty reply raises `MalformedReplyError`
 - `app/api/chat.py`: deflected path uses a turn filter + extracted `_pump` piece loop, one bounded recovery when the filter wiped the reply, `MalformedReplyError` caught before the generic `LLMError` (dedicated error copy, no `done`/`query_log`), log line ends `…retries=N scaffold_stripped=N`
 - `app/rag/scaffolding.py`: `stripped_spans` (allowed by task); tests: 10 new unit (grounded matrix) + 3 new integration (deflected matrix), fakes now feed the filter like the real client
 **Results**
 - `uv run pytest tests/unit/test_agent.py tests/integration/test_chat_api.py -v --no-cov` → 94 passed
 - `uv run pytest --cov=app --cov-report=term-missing` → green, TOTAL **99%** (>90%; all touched files 100%)
 - `uv run ruff check . && uv run pyright` → clean
 - E2E regressions in isolation: `test_harness_aligned_tools.py` (3 passed), `test_chat_rag.py` + `test_agent_document_tools.py` (7 passed)
 **Decisions**
 - Task's message formula conflicted with the locked "single system message" decision — implemented the locked one (replace original system msg with folded correction, history intact), pinned by unit tests
 - Forced-final request is filtered too; scaffolding-only final settles `MalformedReplyError` (no recovery after it — at most one per turn)
 **Next pending task:** `.agent/phases/todo/71_scaffolding_guardrails/04_deflect_prompt.md`
@@ -0,0 +1,77 @@
 ........................................................................ [  4%]
 ........................................................................ [  9%]
 ........................................................................ [ 14%]
 ........................................................................ [ 19%]
 ........................................................................ [ 24%]
 ........................................................................ [ 29%]
 ........................................................................ [ 34%]
 ........................................................................ [ 39%]
 ........................................................................ [ 44%]
 ........................................................................ [ 48%]
 ........................................................................ [ 53%]
 ........................................................................ [ 58%]
 ........................................................................ [ 63%]
 ........................................................................ [ 68%]
 ........................................................................ [ 73%]
 ........................................................................ [ 78%]
 ........................................................................ [ 83%]
 ........................................................................ [ 88%]
 ........................................................................ [ 93%]
 ........................................................................ [ 97%]
 ..............................                                           [100%]
 =============================== warnings summary ===============================
 .venv/lib/python3.13/site-packages/fastapi/testclient.py:1
  /var/home/ducoterra/Projects/Personal/brain_of_reese/.venv/lib/python3.13/site-packages/fastapi/testclient.py:1: StarletteDeprecationWarning: Using `httpx` with `starlette.testclient` is deprecated; install `httpx2` instead.
    from starlette.testclient import TestClient as TestClient  # noqa
 -- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
 ================================ tests coverage ================================
 _______________ coverage: platform linux, python 3.13.13-final-0 _______________
 Name                        Stmts   Miss  Cover
 -----------------------------------------------
 app/__init__.py                 1      0   100%
 app/api/__init__.py             0      0   100%
 app/api/auth.py                22      0   100%
 app/api/chat.py               176      0   100%
 app/api/chats.py              110      0   100%
 app/api/config.py               7      0   100%
 app/api/doc_drafts.py          93      0   100%
 app/api/docs.py                50      0   100%
 app/api/git_sources.py        212      0   100%
 app/api/health.py              10      0   100%
 app/api/steering.py            42      0   100%
 app/api/suggestions.py          8      0   100%
 app/api/sync.py               101      0   100%
 app/config.py                 126      0   100%
 app/core/__init__.py            0      0   100%
 app/core/auth.py               20      0   100%
 app/core/caching.py           108      0   100%
 app/core/debugging.py          29      2    93%
 app/core/docs_push.py          39      0   100%
 app/core/logging.py            13      0   100%
 app/db.py                      21      0   100%
 app/main.py                    52      0   100%
 app/models.py                  86      0   100%
 app/rag/__init__.py             0      0   100%
 app/rag/agent.py              181      0   100%
 app/rag/archive_upload.py     128      0   100%
 app/rag/chunker.py            206      4    98%
 app/rag/git_sources.py         14      0   100%
 app/rag/importer.py           180      3    98%
 app/rag/llm.py                216      0   100%
 app/rag/overview.py            71      0   100%
 app/rag/prompts.py             65      0   100%
 app/rag/retriever.py           94      3    97%
 app/rag/scaffolding.py         55      0   100%
 app/rag/source_removal.py      41      0   100%
 app/rag/sources_meta.py        16      0   100%
 app/rag/suggestions.py         27      0   100%
 app/rag/summarizer.py          24      0   100%
 app/schemas.py                205      0   100%
 -----------------------------------------------
 TOTAL                        2849     12    99%
 coverage gate: app/ 99% (>90%) OK
 All checks passed!
 0 errors, 0 warnings, 0 informations
 validation OK
@@ -0,0 +1,17 @@
 All done. Final report:
 **Task 04 complete — deflection plain-text line (prevention)**
 - `app/rag/prompts.py`: appended "Reply in plain text only — you have no tools in this mode." as its own line to the `DEFLECT_MODE` body (LOW prompt's only change); module + function docstrings carry the phase-71 note (owner-permitted 2026-09-03; marker-keying contract unchanged); stale "LOW byte-identical" claims in the `<tools>` paragraph/comment corrected.
 - `tests/unit/test_prompts.py`: diff/reconstruction pin (LOW = pre-phase text + exactly the one line, `count == 1`, title list untouched), line-present + marker-present pin, no-leak-into-HIGH pin (all 4 HIGH build variants); both empty-steering/overview byte-identity pins updated to the new LOW baseline.
 - `tests/unit/test_agent.py`: LOW byte pin updated to include the line. `tests/unit/test_chat_gate.py`: line pinned present in the deflected `plan_turn` prompt and absent from the HIGH prompt.
 **Results**
 - `uv run pytest tests/unit/test_prompts.py tests/unit/test_chat_gate.py -v --no-cov` → 61 passed
 - `uv run pytest --cov=app --cov-report=term-missing` → 1473 passed, TOTAL 99% (`app/rag/prompts.py` 100%)
 - `uv run ruff check . && uv run pyright` → clean (0 errors)
 - `uv run pytest tests/e2e/test_honest_deflection.py -v --no-cov` → 3 passed (mock marker-keying intact)
 **Decisions**: the line goes on its own line (the completion criterion's "pre-phase text + exactly one new line" diff pin), not appended to the existing sentence. No commit — commits belong to task 05 (phase 70 + 71/01–03 work is still uncommitted in-tree).
 **Next pending task**: `.agent/phases/todo/71_scaffolding_guardrails/05_e2e_commit.md`
@@ -0,0 +1,77 @@
 ........................................................................ [  4%]
 ........................................................................ [  9%]
 ........................................................................ [ 14%]
 ........................................................................ [ 19%]
 ........................................................................ [ 24%]
 ........................................................................ [ 29%]
 ........................................................................ [ 34%]
 ........................................................................ [ 39%]
 ........................................................................ [ 43%]
 ........................................................................ [ 48%]
 ........................................................................ [ 53%]
 ........................................................................ [ 58%]
 ........................................................................ [ 63%]
 ........................................................................ [ 68%]
 ........................................................................ [ 73%]
 ........................................................................ [ 78%]
 ........................................................................ [ 83%]
 ........................................................................ [ 87%]
 ........................................................................ [ 92%]
 ........................................................................ [ 97%]
 .................................                                        [100%]
 =============================== warnings summary ===============================
 .venv/lib/python3.13/site-packages/fastapi/testclient.py:1
  /var/home/ducoterra/Projects/Personal/brain_of_reese/.venv/lib/python3.13/site-packages/fastapi/testclient.py:1: StarletteDeprecationWarning: Using `httpx` with `starlette.testclient` is deprecated; install `httpx2` instead.
    from starlette.testclient import TestClient as TestClient  # noqa
 -- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
 ================================ tests coverage ================================
 _______________ coverage: platform linux, python 3.13.13-final-0 _______________
 Name                        Stmts   Miss  Cover
 -----------------------------------------------
 app/__init__.py                 1      0   100%
 app/api/__init__.py             0      0   100%
 app/api/auth.py                22      0   100%
 app/api/chat.py               176      0   100%
 app/api/chats.py              110      0   100%
 app/api/config.py               7      0   100%
 app/api/doc_drafts.py          93      0   100%
 app/api/docs.py                50      0   100%
 app/api/git_sources.py        212      0   100%
 app/api/health.py              10      0   100%
 app/api/steering.py            42      0   100%
 app/api/suggestions.py          8      0   100%
 app/api/sync.py               101      0   100%
 app/config.py                 126      0   100%
 app/core/__init__.py            0      0   100%
 app/core/auth.py               20      0   100%
 app/core/caching.py           108      0   100%
 app/core/debugging.py          29      2    93%
 app/core/docs_push.py          39      0   100%
 app/core/logging.py            13      0   100%
 app/db.py                      21      0   100%
 app/main.py                    52      0   100%
 app/models.py                  86      0   100%
 app/rag/__init__.py             0      0   100%
 app/rag/agent.py              181      0   100%
 app/rag/archive_upload.py     128      0   100%
 app/rag/chunker.py            206      4    98%
 app/rag/git_sources.py         14      0   100%
 app/rag/importer.py           180      3    98%
 app/rag/llm.py                216      0   100%
 app/rag/overview.py            71      0   100%
 app/rag/prompts.py             65      0   100%
 app/rag/retriever.py           94      3    97%
 app/rag/scaffolding.py         55      0   100%
 app/rag/source_removal.py      41      0   100%
 app/rag/sources_meta.py        16      0   100%
 app/rag/suggestions.py         27      0   100%
 app/rag/summarizer.py          24      0   100%
 app/schemas.py                205      0   100%
 -----------------------------------------------
 TOTAL                        2849     12    99%
 coverage gate: app/ 99% (>90%) OK
 All checks passed!
 0 errors, 0 warnings, 0 informations
 validation OK
@@ -91,6 +91,30 @@ the same budget, and the deflected answer stream goes through
 The per-turn log line records ``retries=N`` after ``total_ms=N`` (0
 when nothing was retried — the field is uniform across all turn
 shapes).
 Tool-scaffolding guardrail (phase 71, deterministic only — owner
 permission 2026-09-03: "deterministic guardrails only right now, forget
 using a model for that"): the raw chat-template tokens
 ``<|tool_call_start|>…<|tool_call_end|>`` the ``lite`` model sometimes
 emits as plain answer text can never reach the user. The DEFLECTED
 path's request runs ``delta.content`` through a caller-owned
 ``ScaffoldingFilter`` (content only — thinking stays raw); when the
 filter wipes the whole reply (visible content 0, ``stripped_chars > 0``)
 the turn gets exactly ONE bounded recovery: the same messages with
 ``agent.CORRECTION_INSTRUCTION`` folded into the single system prompt,
 ``tools=None``, a fresh filter, the same phase-67 retry budget, streamed
 through the same piece loop (extracted as the inner ``_pump`` helper).
 A recovery that also comes back empty — and any grounded round where the
 recovery policy in ``run_agent`` fails — settles with the dedicated
 structured ``error`` frame (``MalformedReplyError`` caught before the
 generic ``LLMError`` handler): no ``done``, no ``query_log`` row,
 byte-for-byte the existing terminal-error shape. The per-turn log line
 records ``scaffold_stripped=N`` after ``retries=N`` — the sum across the
 turn's requests (grounded: the agent's rounds + forced final + any
 recovery, via the holder; deflected: this turn's filters), 0 on clean
 turns (the field is uniform, the phase-67 ``retries=N`` pattern); the
 recovery does not bump ``retries=N`` (it is not a phase-67
 endpoint-retry).
 """
 from __future__ import annotations
@@ -110,7 +134,12 @@ from app.api.steering import load_steering_notes
 from app.config import Settings, get_settings
 from app.db import db_available, get_db
 from app.models import Document, QueryLog
-from app.rag.agent import AgentHolder, run_agent
+from app.rag.agent import (
    CORRECTION_INSTRUCTION,  # phase 71: the harness-owned recovery line
    AgentHolder,
    MalformedReplyError,  # phase 71: the recovery policy's terminal signal
    run_agent,
 )
 from app.rag.llm import (
    EmbeddingError,
    LLMClient,
@@ -123,6 +152,7 @@ from app.rag.llm import (
 from app.rag.overview import load_kb_overview
 from app.rag.prompts import build_deflect_prompt, build_high_prompt
 from app.rag.retriever import RetrievedChunk, retrieve, select_documents, weak_hit_titles
 from app.rag.scaffolding import ScaffoldingFilter  # phase 71: the streaming filter
 from app.rag.suggestions import derive_suggestions
 from app.schemas import (
    ChatDoneEvent,
@@ -373,19 +403,27 @@ async def chat(
            #    ``tools=None`` request anyway (the kill switch).
            holder = AgentHolder()
            answer_stream: AsyncIterator[StreamPiece | ToolCallPiece | RetryPiece]
            deflected_filter: ScaffoldingFilter | None = None
            if plan.deflected:
                # Phase 67: the deflected stream goes through the retry
                # primitive — a dead endpoint is restarted (SSE ``retry``
                # frames) only before its first piece (locked A2); the
                # grounded path stays a plain ``run_agent`` call (task 03
                # makes IT retry internally) — its ``RetryPiece``s flow
-                # through the shared piece loop below.
+                # through the shared piece loop below. Phase 71: the
                # request's content also runs through a caller-owned
                # filter (one per request) — a scaffolding-only reply
                # streams zero ``delta`` frames instead of raw tokens, and
                # the filter's ``stripped_chars`` drives the recovery
                # decision after the piece loop.
                deflected_filter = ScaffoldingFilter()
                answer_stream = chat_stream_retried(
                    llm,
                    messages,
                    tools=None,
                    retries=settings.llm_retries,
                    delay=settings.llm_retry_delay,
                    scaffolding=deflected_filter,
                )
            else:
                answer_stream = run_agent(
@@ -398,8 +436,19 @@ async def chat(
                    holder=holder,
                )
            thinking_chars = 0
-            try:
+            content_chars = 0  # phase 71: the turn's visible (clean) content
-                async for piece in answer_stream:  # StreamPiece | ToolCallPiece | RetryPiece
+            scaffold_stripped = 0  # phase 71: sum across the turn's requests
            async def _pump(
                pieces: AsyncIterator[StreamPiece | ToolCallPiece | RetryPiece],
            ) -> AsyncIterator[str]:
                """One request's piece loop (phase 71 extraction): the
                thinking/tool/retry/delta handling shared by the turn's
                first pass and — deflected path only — the one bounded
                recovery. Behavior-preserving for the first pass (pinned
                by the existing integration suite)."""
                nonlocal thinking_chars, content_chars, retries_used
                async for piece in pieces:  # StreamPiece | ToolCallPiece | RetryPiece
                    if isinstance(piece, ToolCallPiece):
                        # Phase 37 (PLAN §4 extension; phase 70): one SSE
                        # ``tool`` frame per model-requested call.
@@ -435,7 +484,91 @@ async def chat(
                        if settings.stream_thinking:
                            yield sse_event(ChatThinkingEvent(text=piece.text).model_dump())
                    else:
                        content_chars += len(piece.text)
                        yield sse_event({"type": "delta", "text": piece.text})
            try:
                async for frame in _pump(answer_stream):
                    yield frame
                if plan.deflected and deflected_filter is not None:
                    scaffold_stripped = deflected_filter.stripped_chars
                    # Phase 71: the deflected reply's visible content was
                    # wiped by the filter (the scaffolding was the whole
                    # "answer") — the ONE bounded recovery: the same
                    # messages with the correction folded into the single
                    # system prompt, ``tools=None``, a FRESH filter, the
                    # same phase-67 retry budget, streamed through the
                    # same piece loop. A round with real visible content
                    # needs no recovery (the clean content stands).
                    if content_chars == 0 and deflected_filter.stripped_chars > 0:
                        logger.warning(
                            "chat: deflected reply was pure tool-scaffolding "
                            "(%d chars stripped) — running the one bounded "
                            "recovery",
                            deflected_filter.stripped_chars,
                        )
                        recovery_filter = ScaffoldingFilter()
                        recovery_stream = chat_stream_retried(
                            llm,
                            [
                                {
                                    "role": "system",
                                    "content": (
                                        plan.system_prompt + "\n"
                                        + CORRECTION_INSTRUCTION
                                    ),
                                },
                                *messages[1:],
                            ],
                            tools=None,
                            retries=settings.llm_retries,
                            delay=settings.llm_retry_delay,
                            scaffolding=recovery_filter,
                        )
                        async for frame in _pump(recovery_stream):
                            yield frame
                        scaffold_stripped += recovery_filter.stripped_chars
                        if content_chars == 0:
                            # The second empty reply is terminal (at most
                            # one recovery per turn) — the dedicated error
                            # frame below (no done, no query_log row).
                            logger.warning(
                                "chat: the recovery reply was still empty "
                                "(scaffold_stripped=%d) — settling with a "
                                "malformed-reply error",
                                scaffold_stripped,
                            )
                            raise MalformedReplyError(
                                "the deflected model answered in raw "
                                "tool-scaffolding twice in a row — no "
                                "clean answer to stream"
                            )
                else:
                    # Grounded turns: the agent's rounds + forced final +
                    # any recovery already accumulated the turn total on
                    # the holder (the deflected fallback is 0 — the agent
                    # never runs, so this branch is grounded-only).
                    scaffold_stripped = holder.scaffold_stripped
            except MalformedReplyError as e:
                # Phase 71: the recovery policy's terminal signal —
                # caught BEFORE the generic LLMError handler (it
                # subclasses it), so the dedicated copy reaches the UI;
                # the generic "dropped the connection" copy stays for
                # transport failures.
                logger.error(
                    "chat: malformed reply after the one bounded recovery "
                    "question=%r total_ms=%d — %s",
                    request.message,
                    int((time.monotonic() - started) * 1000),
                    e,
                )
                settled = True  # terminal: the error frame settles the turn
                yield sse_event(
                    ChatErrorEvent(
                        detail="The model returned a malformed reply — please try again."
                    ).model_dump()
                )
                return
            except LLMError as e:
                logger.error(
                    "chat: LLM stream failed question=%r total_ms=%d — %s",
@@ -504,7 +637,7 @@ async def chat(
            logger.info(
                "question=%r embed_ms=%d top_score=%.3f fts_hits=%d summary_hits=%d tuning=%d "
                "kb_chars=%d threshold=%.2f deflected=%s sources=%r thinking_chars=%d "
-                "tool_calls=%d total_ms=%d retries=%d",
+                "tool_calls=%d total_ms=%d retries=%d scaffold_stripped=%d",
                request.message,
                embed_ms,
                plan.top_score,
@@ -519,6 +652,7 @@ async def chat(
                holder.tool_calls,
                total_ms,
                retries_used,
                scaffold_stripped,
            )
            settled = True  # terminal: the done frame settles the turn
            yield sse_event(
@@ -95,6 +95,27 @@ task 04):
   exactly one round. With ``settings.llm_retries=0`` every request is a
   single plain attempt (the pre-phase-67 path).
 Scaffolding guardrail (phase 71, deterministic only — owner permission
 2026-09-03: "deterministic guardrails only right now, forget using a
 model for that"): every model request (each round, the forced final,
 and any recovery) runs its ``delta.content`` through a fresh caller-
 owned :class:`app.rag.scaffolding.ScaffoldingFilter`, so raw
 ``<|tool_call_start|>…<|tool_call_end|>`` tokens can never reach the
 user as answer text. A round that ends with NO visible content AND a
 non-empty strip (the scaffolding was the whole "answer") gets exactly
 ONE bounded recovery: one extra request with ``tools=None``, the same
 messages with :data:`CORRECTION_INSTRUCTION` folded into the original
 single system message, a fresh filter, and the same phase-67 retry
 budget. A recovery that also comes back empty — or a round with no
 strip and no content (today's empty/thinking-only answer) — settles as
 before; a second empty reply raises :class:`MalformedReplyError` (the
 API layer turns it into the dedicated error frame). A round with real
 visible content plus scaffolding needs no recovery (the clean content
 stands), and a scaffolding-only round that also carried tool calls
 needs none either (the tool ran) — the policy keys on the no-calls
 exit only. No model participates in detection or repair: the
 registry + the fixed retry policy are the whole guardrail.
 The DB accessors (:func:`list_catalog`, :func:`list_source_names`,
 :func:`find_document`, :func:`all_documents`) and the
 :func:`grep_document` line matcher are module-level functions so unit
@@ -116,11 +137,13 @@ from app.models import Document
 from app.rag.git_sources import effective_sources
 from app.rag.llm import (
    LLMClient,
    LLMError,
    RetryPiece,
    StreamPiece,
    ToolCallPiece,
    chat_stream_retried,
 )
 from app.rag.scaffolding import ScaffoldingFilter
 from app.rag.source_removal import resolve_source_name
 logger = logging.getLogger("app.agent")
@@ -228,6 +251,33 @@ UNKNOWN_TOOL = "Unknown tool."
 MISSING_READ_ARGS = "read requires a string argument 'path'."
 MISSING_SEARCH_ARGS = "grep requires a string argument 'pattern'."
 #: The harness-owned recovery line (phase 71, task 03) — folded into the
 #: ORIGINAL single system message of the one bounded recovery request
 #: (``system_prompt + "\n" + CORRECTION_INSTRUCTION``; provider-safe,
 #: the user message stays last). Verbatim constant: the E2E mock
 #: (task 05) keys on a stable substring of it, so it must not drift.
 CORRECTION_INSTRUCTION: str = (
    "Your previous reply contained raw tool-call markup, which is not "
    "interpreted here. Answer the user's question directly in plain "
    "text — no tool syntax."
 )
 class MalformedReplyError(LLMError):
    """The model kept replying in raw tool-scaffolding (phase 71).
    Raised ONLY by the recovery policy — :func:`run_agent` (grounded
    path) and ``app.api.chat`` (deflected path) — when the one bounded
    ``tools=None`` recovery still comes back with no visible content.
    It is never raised from inside a stream, so
    :func:`app.rag.llm.chat_stream_retried`'s retry-before-first-piece
    rule never sees it. The API layer catches it BEFORE the generic
    :class:`LLMError` handler and settles the turn with the dedicated
    "malformed reply" error frame (no ``done``, no ``query_log`` row).
    Deterministic only (owner permission 2026-09-03): no model
    participates in detection or repair.
    """
 #: Search caps (owner-locked A5, phase 68): a global per-call match cap
 #: (across documents, in catalog order) and a per-match-line char limit.
 SEARCH_MAX_MATCHES = 20
@@ -352,10 +402,16 @@ class AgentHolder:
    rejected calls (unknown tool, unknown/missing arguments or document,
    already-in-context) do not count. Drives the per-turn log line's
    ``tool_calls=N`` field (task 04).
    ``scaffold_stripped``: how many chars of tool-scaffolding the
    turn's filters removed across the turn's requests (rounds + the
    forced final + any recovery, phase 71) — drives the per-turn log
    line's ``scaffold_stripped=N`` field on grounded turns (the
    deflected path computes its own total in ``app.api.chat``).
    """
    read_docs: list[Document] = field(default_factory=list)
    tool_calls: int = 0
    scaffold_stripped: int = 0
 def _execute_tool(
@@ -476,6 +532,24 @@ async def run_agent(
    ``settings.llm_retry_delay``); a round that already streamed pieces
    fails the turn as before.
    Scaffolding recovery (phase 71, deterministic only): every request —
    each round, the forced final, and any recovery — runs its content
    through a fresh :class:`app.rag.scaffolding.ScaffoldingFilter`. A
    round that ends with NO visible content but a non-empty strip (the
    scaffolding was the whole "answer") gets exactly ONE recovery:
    ``tools=None``, :data:`CORRECTION_INSTRUCTION` folded into the
    original single system message (the rest of the history — user
    message and tool results — unchanged), a fresh filter, the same
    retry budget. A clean recovery ends the turn; a second empty reply
    raises :class:`MalformedReplyError` (terminal — the API layer turns
    it into the dedicated error frame). A round with visible content
    plus scaffolding needs no recovery (the clean content stands), and
    a scaffolding-only round that also carried tool calls needs none
    (the tool ran) — the policy keys on the no-calls exit only. The
    per-span strip warning log (each span truncated to 200 chars) is
    the capture mechanism for new registry entries; *holder* accumulates
    the turn's ``scaffold_stripped`` total for the API layer's log line.
    ``seed_docs`` are the documents the retrieval already put in context
    (they shape the *system_prompt* the caller built); re-reading one of
    them is rejected as "Already in your context." — the rejection counts
@@ -494,6 +568,12 @@ async def run_agent(
    rounds = 0
    while True:
        calls: list[ToolCallPiece] = []
        # Phase 71: one fresh filter per round (one per model request —
        # the retry attempts of this logical request share it: a restart
        # only happens while the filter was never fed). Content-only:
        # thinking pieces pass through raw.
        round_filter = ScaffoldingFilter()
        round_content = 0  # visible (clean) content chars this round
        # Phase 48: bind the round's stream so a consumer abandon
        # (GeneratorExit into the yield below) tears down the in-flight
        # model stream deterministically — not GC-dependent. Phase 67:
@@ -511,16 +591,97 @@ async def run_agent(
            tools=tools,
            retries=settings.llm_retries,
            delay=settings.llm_retry_delay,
            scaffolding=round_filter,
        )
        try:
            async for piece in stream:
                if isinstance(piece, ToolCallPiece):
                    calls.append(piece)
                elif isinstance(piece, StreamPiece) and piece.kind == "content":
                    round_content += len(piece.text)
                yield piece
        finally:
            await stream.aclose()
        # Phase 71: the per-strip-event capture log — one warning per
        # stripped span, truncated to 200 chars (how a new scaffolding
        # format gets captured and added to the registry) — and the turn
        # total for the API layer's ``scaffold_stripped=N`` log field.
        holder.scaffold_stripped += round_filter.stripped_chars
        for span in round_filter.stripped_spans:
            logger.warning(
                "agent: stripped %d chars of tool-scaffolding in round %d: %r",
                len(span),
                rounds + 1,
                span[:200],
            )
        if not calls:
-            return  # the answer was streamed
+            if round_content > 0:
                return  # the answer was streamed (the clean content stands)
            if round_filter.stripped_chars == 0:
                # Empty/thinking-only answer — today's behavior, unchanged
                # (the UI handles it); the guardrail keys on a strip.
                return
            # Phase 71: the scaffolding was the whole "answer" — the ONE
            # bounded recovery (a fixed policy, not a conversation):
            # ``tools=None``, the correction folded into the ORIGINAL
            # single system message (provider-safe — the user message and
            # any tool history stay in place), a fresh filter, the same
            # phase-67 retry budget.
            logger.warning(
                "agent: round %d was pure tool-scaffolding (%d chars stripped) "
                "— running the one bounded recovery",
                rounds + 1,
                round_filter.stripped_chars,
            )
            messages_recovered = [
                {
                    "role": "system",
                    "content": system_prompt + "\n" + CORRECTION_INSTRUCTION,
                },
                *messages[1:],
            ]
            recovery_filter = ScaffoldingFilter()
            recovered = chat_stream_retried(
                llm,
                cast("list[dict[str, str]]", messages_recovered),
                tools=None,
                retries=settings.llm_retries,
                delay=settings.llm_retry_delay,
                scaffolding=recovery_filter,
            )
            recovery_content = 0
            try:
                async for piece in recovered:
                    if isinstance(piece, StreamPiece) and piece.kind == "content":
                        recovery_content += len(piece.text)
                    yield piece
            finally:
                await recovered.aclose()
            holder.scaffold_stripped += recovery_filter.stripped_chars
            for span in recovery_filter.stripped_spans:
                logger.warning(
                    "agent: stripped %d chars of tool-scaffolding in the "
                    "recovery after round %d: %r",
                    len(span),
                    rounds + 1,
                    span[:200],
                )
            if recovery_content > 0:
                return  # the recovery answered — the turn ends
            # The second empty reply is terminal (at most one recovery per
            # turn). Raised OUTSIDE the stream, so chat_stream_retried's
            # retry rule never sees it; the API layer catches it before
            # the generic LLMError handler.
            logger.warning(
                "agent: the recovery reply was still empty "
                "(scaffold_stripped=%d) — settling with a malformed-reply error",
                holder.scaffold_stripped,
            )
            raise MalformedReplyError(
                f"the model answered in raw tool-scaffolding twice in a row "
                f"(round {rounds + 1} plus one recovery) — no clean answer "
                "to stream"
            )
        call = calls[0]  # a stream can carry several calls; run the first
        result = _execute_tool(db, call, seed_docs, holder)
        rounds += 1  # every call the model emits consumes a round
@@ -558,17 +719,46 @@ async def run_agent(
            # teardown as the loop rounds (consumer abandon mid-final
            # answer must still close the model's stream). Phase 67: the
            # forced call retries under the same locked-A2 rule as the
-            # loop rounds.
+            # loop rounds. Phase 71: the forced final runs through a
            # fresh filter too — raw scaffolding can never reach the
            # user from ANY grounded request.
            final_filter = ScaffoldingFilter()
            final = chat_stream_retried(
                llm,
                cast("list[dict[str, str]]", messages),
                tools=None,
                retries=settings.llm_retries,
                delay=settings.llm_retry_delay,
                scaffolding=final_filter,
            )
            final_content = 0
            try:
                async for piece in final:
                    if isinstance(piece, StreamPiece) and piece.kind == "content":
                        final_content += len(piece.text)
                    yield piece
            finally:
                await final.aclose()
            # Phase 71: the same capture log + turn total; a
            # scaffolding-only forced final (this turn used no recovery,
            # so nothing is doubled up) settles with the same terminal
            # malformed-reply error rather than a silently empty answer.
            holder.scaffold_stripped += final_filter.stripped_chars
            for span in final_filter.stripped_spans:
                logger.warning(
                    "agent: stripped %d chars of tool-scaffolding in the "
                    "forced final answer (round %d): %r",
                    len(span),
                    rounds + 1,
                    span[:200],
                )
            if final_content == 0 and final_filter.stripped_chars > 0:
                logger.warning(
                    "agent: the forced final answer was pure tool-scaffolding "
                    "— settling with a malformed-reply error"
                )
                raise MalformedReplyError(
                    "the forced final answer was raw tool-scaffolding — no "
                    "clean answer to stream"
                )
            return
@@ -23,13 +23,18 @@ import json
 import logging
 from collections.abc import AsyncGenerator
 from dataclasses import dataclass
-from typing import Any, Literal, cast
+from typing import TYPE_CHECKING, Any, Literal, cast
 from openai import AsyncOpenAI, AsyncStream
 from openai.types.chat import ChatCompletionChunk, ChatCompletionMessageParam
 from app.config import Settings, get_settings
 if TYPE_CHECKING:
    # Phase 71: the filter type is only needed for typing (the module
    # stays import-graph-clean; callers pass their own instances).
    from app.rag.scaffolding import ScaffoldingFilter
 logger = logging.getLogger("app.llm")
@@ -338,6 +343,7 @@ class LLMClient:
        self,
        messages: list[dict[str, str]],
        tools: list[dict[str, Any]] | None = None,
        scaffolding: ScaffoldingFilter | None = None,
    ) -> AsyncGenerator[StreamPiece | ToolCallPiece, None]:
        """Stream assistant pieces from the chat model (PLAN A5/A15, phase 17).
@@ -365,13 +371,26 @@ class LLMClient:
        ``id`` and ``function.name`` on the first partial and
        ``function.arguments`` in fragments — which are accumulated into
        one :class:`ToolCallPiece` per call, yielded in index order at
-        stream end (or immediately once a chunk carries
+        stream end (after the stream's chunks are exhausted — aipi ends
-        ``finish_reason="tool_calls"``). Malformed ``arguments`` JSON
+        the stream at ``finish_reason="tool_calls"``, so this is the
-        raises :class:`LLMError`. Wire convention verified live against
+        wire's emission point). Malformed ``arguments`` JSON raises
        :class:`LLMError`. Wire convention verified live against
        aipi's ``turbo`` on 2026-08-26 via
        ``uv run python -m scripts.llm_probe --tools`` (phase 37, task 01:
        ``probe: turbo tool_calls=supported 2026-08-26``).
        Scaffolding guardrail (phase 71): the caller may pass a
        ``ScaffoldingFilter`` — one per request, caller-owned (this
        method never creates or resets one). When present, only
        **content** is filtered: ``delta.content`` is fed through the
        filter and only the clean text is yielded (an empty clean
        result yields **no** piece — no empty ``delta`` frames); thinking
        pieces are never filtered (the scratchpad stays raw, phase 17).
        At stream end the filter's held tail is flushed to a content
        piece **before** any tool-call materialization (content-
        before-tools wire convention). ``None`` (the default) keeps
        today's byte-identical raw path for callers that opt out.
        Any failure (network, HTTP, malformed stream) surfaces as
        :class:`LLMError` so the API layer can turn it into an SSE
        ``error`` event instead of a hung request.
@@ -405,7 +424,6 @@ class LLMClient:
                await self._client.chat.completions.create(**kwargs),
            )
            calls: dict[int, _ToolCallSlot] = {}
            emitted = False
            async for chunk in stream:
                if not chunk.choices:
                    continue
@@ -436,16 +454,22 @@ class LLMClient:
                    yield StreamPiece("thinking", reasoning)
                content = delta.content
                if content:
-                    yield StreamPiece("content", content)
+                    if scaffolding is not None:
-                if (
+                        # Phase 71: content only — an empty clean result
-                    calls
+                        # yields nothing (no empty delta frames).
-                    and not emitted
+                        cleaned = scaffolding.feed(content)
-                    and getattr(choice, "finish_reason", None) == "tool_calls"
+                        if cleaned:
-                ):
+                            yield StreamPiece("content", cleaned)
-                    for piece in _materialize_tool_calls(calls):
+                    else:
-                        yield piece
+                        yield StreamPiece("content", content)
-                    emitted = True
+            # Phase 71: flush the filter's held tail at stream end, BEFORE
-            if calls and not emitted:
+            # any tool-call materialization — flushed-tail content precedes
            # ToolCallPieces (content-before-tools wire convention).
            if scaffolding is not None:
                tail = scaffolding.flush()
                if tail:
                    yield StreamPiece("content", tail)
            if calls:
                for piece in _materialize_tool_calls(calls):
                    yield piece
        except LLMError:
@@ -469,6 +493,7 @@ async def chat_stream_retried(
    tools: list[dict[str, Any]] | None = None,
    retries: int = 0,
    delay: float = 0.0,
    scaffolding: ScaffoldingFilter | None = None,
 ) -> AsyncGenerator[StreamPiece | ToolCallPiece | RetryPiece, None]:
    """Stream a chat turn, retrying a dead endpoint (phase 67).
@@ -493,6 +518,13 @@ async def chat_stream_retried(
    The request is restarted byte-identical: ``chat_stream`` is stateless,
    so every attempt is opened with the SAME *messages*/*tools*.
    Scaffolding guardrail (phase 71): *scaffolding* is passed through to
    every attempt's ``chat_stream``. The SAME caller-owned filter object
    across the retry attempts of one logical request is safe by
    construction: a restarted attempt only happens while no piece was
    emitted, i.e. the filter was never fed (its pending buffer is still
    empty).
    Teardown (phase 48, extended): every attempt's stream is explicitly
    closed in a ``finally`` — normal exhaustion, a terminal
    :class:`LLMError`, and a consumer abandon (``GeneratorExit`` mid-attempt
@@ -502,7 +534,7 @@ async def chat_stream_retried(
    max_attempts = retries + 1
    for attempt in range(1, max_attempts + 1):
        emitted = False
-        stream = llm.chat_stream(messages, tools=tools)
+        stream = llm.chat_stream(messages, tools=tools, scaffolding=scaffolding)
        try:
            async for piece in stream:
                emitted = True
@@ -30,7 +30,19 @@ Agent tools (phase 37; phase 70: the copy teaches the harness-aligned
 may extend its context through the three server-side tools (round-
 capped, see :mod:`app.rag.agent`; the cap is the bound and this section
 does not re-state it, phase 45). The LOW/deflection prompt never
-carries it and stays byte-identical to the pre-phase text.
+carries it (phase 71: the LOW prompt's only addition is the
 plain-text line below — it still has no ``<tools>`` section).
 Deflection plain-text line (phase 71, owner-permitted 2026-09-03):
 the otherwise-locked ``LOW`` prompt gains exactly one instruction
 line — "Reply in plain text only — you have no tools in this mode."
 — appended to the ``DEFLECT_MODE`` body: a deflected turn offers no
 tools, so any tool markup there is always wrong, and the line closes
 the door at the prompt (the deterministic filter + one bounded
 recovery in :mod:`app.rag.scaffolding` / :mod:`app.rag.agent` is the
 backstop). The ``DEFLECT_MODE`` marker and everything else in the
 prompt stay put — the E2E mock LLM keys on the marker's *presence*,
 not the wording, so that contract is unchanged.
 """
 from __future__ import annotations
@@ -83,9 +95,10 @@ _KB_INTRO = (
 #: not re-state it, phase 45). Appended after the mode body
 #: (``<documents>``), so the instructions are the last thing the model
 #: reads. The LOW/deflection prompt never carries it — a deflection has
-#: no grounded context to extend — and stays byte-identical to the
+#: no grounded context to extend (phase 71: the LOW prompt's only
-#: pre-phase text. The E2E mock keys off the ``<tools>`` marker's
+#: addition is the plain-text line in :func:`build_deflect_prompt`).
-#: *presence*, not this wording.
+#: The E2E mock keys off the ``<tools>`` marker's *presence*, not this
 #: wording.
 TOOLS_SECTION: str = (
    "<tools>\n"
    "You may extend your context with three tools. `ls` lists the "
@@ -224,7 +237,10 @@ def build_deflect_prompt(
    Section order (phase 31): ``<relevance>`` → ``<knowledge_base>`` →
    ``<tuning>`` → ``DEFLECT_MODE`` body; empty steering/overview omit
-    their section, keeping the prompt byte-identical to the pre-phase text.
+    their section, keeping the prompt byte-identical to the pre-phase
    text. The body ends with the phase-71 plain-text line (owner-
    permitted 2026-09-03 — the LOW prompt's only change): a deflected
    turn offers no tools, so any tool markup there is always wrong.
    """
    weak = "\n".join(f"- {t}" for t in titles) if titles else "(nothing close at all)"
    mid = "\n".join(
@@ -239,5 +255,9 @@ def build_deflect_prompt(
        + "DEFLECT_MODE: retrieval was weak — the titles below are the closest "
        "your notes come to the question. They are titles only; do not pretend "
        "they answer it. Use them to propose 2-3 alternative questions.\n"
        # Phase 71 (owner-permitted 2026-09-03): the one plain-text line
        # — prevention at the prompt. The E2E mock keys on the
        # DEFLECT_MODE marker's presence, so appending is safe.
        "Reply in plain text only — you have no tools in this mode.\n"
        + weak
    )
@@ -0,0 +1,170 @@
 """Deterministic tool-scaffolding guardrail — pattern registry + streaming
 filter (phase 71, task 01).
 The ``lite`` chat model occasionally emits its own chat-template
 tool-scaffolding as plain answer text (incident 2026-09-03: a deflected
 round streamed ``<|tool_call_start|>[read(path='…')]<|tool_call_end|>``
 into the UI although no tools were offered). This module is the
 detection half of the guardrail: a fixed registry of *observed*
 scaffolding forms and a streaming state machine that strips them from
 ``delta.content`` as it flows. Pure Python — no I/O, no logging, no
 model: the strip *warning* log and the one bounded recovery are emitted
 by the integration layer (phase 71, tasks 02–03), so this module stays
 unit-testable in isolation.
 Content only: thinking pieces are the model's raw reasoning by design
 (phase 17) and are never filtered — the guardrail protects the answer,
 not the scratchpad.
 """
 import re
 __all__ = ["SCAFFOLD_PATTERNS", "ScaffoldingFilter"]
 #: The known scaffolding forms. The registry is the extension point: a
 #: new entry needs an observed capture (the strip warning log, task 03)
 #: + a unit fixture in ``tests/unit/test_scaffolding_filter.py`` — no
 #: speculative entries. Every entry here traces to the 2026-09-03
 #: incident (the span form is the deflected round's raw text; the
 #: standalone siblings are from the same tokenizer family).
 SCAFFOLD_PATTERNS: tuple[re.Pattern, ...] = (
    # The observed span (incident 2026-09-03) — non-greedy, so multiple
    # spans in one buffer each strip to their own end token.
    re.compile(r"<\|tool_call_start\|>[\s\S]*?<\|tool_call_end\|>"),
    # Standalone sibling token (same tokenizer family).
    re.compile(r"<\|tool_calls\|>"),
    # Standalone sibling token (same tokenizer family).
    re.compile(r"<\|tool_call\|>"),
 )
 _SPAN_START = "<|tool_call_start|>"
 _SPAN_END = "<|tool_call_end|>"
 #: The literal openings a live (possibly incomplete) match can begin with.
 _OPENINGS: tuple[str, ...] = (_SPAN_START, "<|tool_calls|>", "<|tool_call|>")
 #: Boundedness bound: in NORMAL state the held tail is at most one char
 #: short of the longest opening (an open span is unbounded, but it is
 #: being *stripped*, never emitted).
 _MAX_OPENING = max(len(opening) for opening in _OPENINGS)
 def _leftmost_match(buf: str) -> re.Match | None:
    """The leftmost complete match among :data:`SCAFFOLD_PATTERNS`, or
    None when the buffer holds no complete scaffolding form.
    (Two patterns cannot match at the same start position — their
    literals diverge after ``<|tool_call`` — so the leftmost start is
    unambiguous.)
    """
    best: re.Match | None = None
    for pattern in SCAFFOLD_PATTERNS:
        match = pattern.search(buf)
        if match is not None and (best is None or match.start() < best.start()):
            best = match
    return best
 def _hold_index(buf: str) -> int:
    """Index where the live hold-tail begins; everything from it stays
    pending (called only after the strip loop found no complete match).
    The tail is the leftmost of two candidates: an **open span** (a
    start token with no end token after it — the span may continue in
    future chunks, so everything from that start token onward is held)
    or a **live prefix** (the longest suffix that is a proper prefix of
    any opening literal — the token may complete in future chunks).
    Everything before the leftmost candidate is safe to emit: no
    complete match remains, and no match can grow from earlier text,
    since a complete opening earlier in the buffer would already have
    matched (span) or be a candidate in its own right (standalone).
    """
    if not buf:
        return 0
    hold = len(buf)
    start = buf.find(_SPAN_START)
    if start != -1 and buf.find(_SPAN_END, start + 1) == -1:
        hold = start
    for k in range(min(len(buf), _MAX_OPENING - 1), 0, -1):
        suffix = buf[-k:]
        if any(k < len(opening) and opening.startswith(suffix) for opening in _OPENINGS):
            hold = min(hold, len(buf) - k)
            break
    return hold
 class ScaffoldingFilter:
    """Streaming state machine over a model **content** stream (phase 71).
    One instance per model request (callers create fresh instances —
    the phase-67 retry-after-dead-attempt case is safe: a dead attempt
    never fed the filter). Feed the raw ``delta.content`` chunks; each
    :meth:`feed` returns the clean text safe to emit *now*; :meth:`flush`
    emits the tail at stream end.
    ``feed`` appends the chunk to the internal pending buffer, then
    (a) repeatedly takes the leftmost complete match among
    :data:`SCAFFOLD_PATTERNS` — drops it (counting it into
    :attr:`stripped_chars`) and continues — until none remains; then
    (b) checks the buffer tail for a **live prefix** via
    :func:`_hold_index`: the longest suffix that is a proper prefix of
    any opening literal, or an **open span** (a start token with no end
    token yet — everything from that start token onward is held, since
    the span may continue in future chunks). Everything before the held
    tail is emitted; the held tail becomes the new pending state.
    Bounded: the held tail in NORMAL state is ≤ the longest opening
    token minus one char; in an open span it is unbounded but is being
    *stripped*, never emitted.
    ``flush`` (end of stream) emits the pending tail **as-is**: a
    partial marker at EOF is content, not scaffolding — a documented,
    pinned choice (a stream that ends mid-``<|tool_call_st`` must not
    be silently eaten, and a lone ``<|tool_call_end|>`` without a start
    is prose the user sent or the model produced outside a span).
    The stripped spans themselves are exposed read-only
    (:attr:`stripped_spans`) so the integration layer can log one
    warning per strip event (phase 71 task 03) — that log line is how
    a *new* scaffolding format gets captured and added to the registry.
    """
    def __init__(self) -> None:
        self._pending = ""
        self._stripped_chars = 0
        self._stripped_spans: list[str] = []
    @property
    def stripped_chars(self) -> int:
        """Total characters removed so far (read by the caller after the
        round/turn — the strip warning log, phase 71 task 03)."""
        return self._stripped_chars
    @property
    def stripped_spans(self) -> list[str]:
        """The raw spans removed so far, in stream order (one entry per
        strip event — the integration layer's per-span warning log
        truncates each to 200 chars). Read-only: callers must not
        mutate the filter's state."""
        return list(self._stripped_spans)
    def feed(self, chunk: str) -> str:
        """Append *chunk* to the pending buffer; return the clean text
        safe to emit now (possibly ``""`` — e.g. while an open span or a
        partial marker is still held). An empty chunk is a no-op."""
        if not chunk:
            return ""
        buf = self._strip_complete(self._pending + chunk)
        hold = _hold_index(buf)
        self._pending = buf[hold:]
        return buf[:hold]
    def flush(self) -> str:
        """End of stream: emit the pending tail as-is and reset it."""
        tail = self._pending
        self._pending = ""
        return tail
    def _strip_complete(self, buf: str) -> str:
        """Drop leftmost-complete matches until none remains (step a)."""
        while (match := _leftmost_match(buf)) is not None:
            self._stripped_chars += match.end() - match.start()
            self._stripped_spans.append(match.group(0))
            buf = buf[: match.start()] + buf[match.end():]
        return buf
@@ -127,6 +127,35 @@ Implements just enough of the aipi surface:
    specific phrase — same convention as ``think in paragraphs``); no
    existing E2E question or fixture file contains the trigger, so
    every other suite is unaffected.
  - user message containing ``emit raw tool markup``
    (``SCAFFOLD_TRIGGER``, phase 71, tool-scaffolding guardrails — the
    2026-09-03 incident where a deflected round streamed the model's
    raw ``<|tool_call_start|>…<|tool_call_end|>`` markup into the UI)
    **or** ``always emit raw tool markup``
    (``SCAFFOLD_ALWAYS_TRIGGER``, checked FIRST — it contains the
    former phrase) -> the deterministic SCAFFOLDING flow, independent
    of the ``<tools>`` marker (both grounded and deflected turns hit
    it):
      * ``SCAFFOLD_ALWAYS_TRIGGER``: EVERY request (the one bounded
        recovery included) streams ONLY ``delta.content`` chunks
        carrying the incident span ``SCAFFOLD_SPAN`` —
        ``<|tool_call_start|>[read(path='search_docs/reese-notes.md')]
        <|tool_call_end|>`` — split across the mock's 12-char chunks
        (the filter's boundary path), ``finish_reason: "stop"``, no
        structured ``tool_calls``, no reasoning — the terminal
        malformed-reply path.
      * ``SCAFFOLD_TRIGGER``: request 1 (no ``CORRECTION_INSTRUCTION``
        in the system prompt) streams the same scaffolding-only span;
        request 2 (the system prompt carries the harness constant — a
        stable substring of ``app.rag.agent.CORRECTION_INSTRUCTION``,
        IMPORTED into this module so the mock can never drift from
        it: the one bounded recovery, ``tools=None`` with the
        correction folded into the single system prompt) streams the
        clean ``SCAFFOLD_RECOVERY_ANSWER`` — the recovery path.
    Checked BEFORE the ``SEARCH_TRIGGER`` / ``TOOLS_TRIGGER`` flows
    (the trigger needs no ``<tools>`` section); no existing E2E
    question or fixture file contains the phrase, so every other
    suite is unaffected.
  - user message containing ``show me a table`` (phase 44, markdown
    tables, TODO.md L6) -> the fixed table answer (``TABLE_ANSWER``):
    a 3-column service table, an ``<img onerror>`` XSS probe line, and
@@ -188,6 +217,8 @@ from typing import Any
 from fastapi import FastAPI
 from fastapi.responses import JSONResponse, StreamingResponse
 from app.rag.agent import CORRECTION_INSTRUCTION  # phase 71: the harness constant
 app = FastAPI()
 DIM = 768
@@ -364,6 +395,51 @@ ALWAYS_FAIL_TRIGGER = "always fail"
 #: bag-of-words vector — the endpoint's pre-stream embedding retry loop.
 EMBED_FAIL_TRIGGER = "embed fail once"
 # ---------------------------------------------------------------------------
 # Phase 71 (tool-scaffolding guardrails, 2026-09-03 incident):
 # deterministic raw-markup flows — see the module docstring
 # ---------------------------------------------------------------------------
 #: A user message containing this substring (case-insensitive) drives
 #: the scaffolding flow: request 1 streams ONLY the incident's raw tool
 #: markup as ``delta.content``; the follow-up request carrying the
 #: harness correction in the system prompt (the one bounded recovery)
 #: streams the clean answer. Independent of the ``<tools>`` marker —
 #: both grounded and deflected turns hit it. Existing E2E questions do
 #: not contain the phrase, so every other suite is unaffected.
 SCAFFOLD_TRIGGER = "emit raw tool markup"
 #: A user message containing this substring (checked BEFORE
 #: ``SCAFFOLD_TRIGGER`` — it contains that phrase) streams the
 #: scaffolding-only span on EVERY request, recovery included — the
 #: terminal malformed-reply path (the dedicated error frame, no done).
 SCAFFOLD_ALWAYS_TRIGGER = "always emit raw tool markup"
 #: The incident span (2026-09-03): the model's chat-template tool
 #: syntax, emitted as plain ``delta.content`` although no tools were
 #: offered. Streamed through the mock's 12-char chunking, so it always
 #: spans ≥2 wire chunks (the filter's boundary path).
 SCAFFOLD_SPAN = (
    "<|tool_call_start|>[read(path='search_docs/reese-notes.md')]"
    "<|tool_call_end|>"
 )
 #: The clean answer the one bounded recovery produces (byte-stable —
 #: the dedicated E2E suite asserts the recovered bubble and the wire's
 #: delta text against it).
 SCAFFOLD_RECOVERY_ANSWER = "Here is the plain-text answer the recovery produced."
 #: The stable substring of the harness-owned correction constant the
 #: recovery request carries in its system prompt. Keyed on a substring
 #: (not the whole constant) so a re-wrap of the constant cannot silently
 #: re-route the mock; the module-level assert below fails loudly if the
 #: substring ever leaves the constant (the mock must never drift from
 #: ``app.rag.agent.CORRECTION_INSTRUCTION``).
 _CORRECTION_MARKER = "no tool syntax"
 assert _CORRECTION_MARKER in CORRECTION_INSTRUCTION, (
    "mock drift: the correction marker left CORRECTION_INSTRUCTION"
 )
 #: One DEAD app-level chat attempt costs exactly this many HTTP POSTs
 #: while the endpoint stays down: the openai SDK's default policy
 #: (max_retries=2 — the app's ``LLMClient`` keeps it) re-POSTs a 500'd
@@ -535,6 +611,35 @@ def _search_flow(body: dict[str, Any]) -> tuple[str, ...] | None:
    return ("search",)
 def _scaffold_flow(body: dict[str, Any]) -> str | None:
    """Classify a phase-71 scaffolding request (see the module docstring).
    * ``"scaffold"`` — stream ONLY the incident span
      (``SCAFFOLD_SPAN``) as ``delta.content`` chunks: ``finish_reason:
      "stop"``, no structured ``tool_calls``, no reasoning. EVERY
      request for ``SCAFFOLD_ALWAYS_TRIGGER`` (the recovery included),
      and the FIRST request of ``SCAFFOLD_TRIGGER`` (no correction in
      the system prompt yet).
    * ``"recovery"`` — ``SCAFFOLD_TRIGGER`` whose system prompt carries
      the harness correction (the one bounded recovery: ``tools=None``,
      the constant folded into the single system prompt by
      ``app.api.chat`` / ``app.rag.agent``): stream the clean
      ``SCAFFOLD_RECOVERY_ANSWER``.
    * ``None`` — not the scaffolding flow. The discrimination is
      stateless, like the other marker flows: the trigger phrase in
      the user message plus the correction's presence in the system
      prompt.
    """
    user = _user(body).lower()
    if SCAFFOLD_ALWAYS_TRIGGER in user:  # checked FIRST — it contains SCAFFOLD_TRIGGER
        return "scaffold"
    if SCAFFOLD_TRIGGER in user:
        if _CORRECTION_MARKER in _system(body):
            return "recovery"
        return "scaffold"
    return None
 def _tool_flow(body: dict[str, Any]) -> tuple[str, ...] | None:
    """Classify a marker request into one step of the tool flow.
@@ -1061,6 +1166,29 @@ def chat_completions(body: dict[str, Any]) -> Any:
            if _chat_dead(RETRY_TRIGGER, RETRY_DEAD_ATTEMPTS):
                return _llm_500(RETRY_TRIGGER)
            _fail_posts[RETRY_TRIGGER] = 0  # the answer streamed — restart
        # Phase 71 (tool-scaffolding guardrails): the deterministic raw-
        # markup flow — checked BEFORE the search/tool marker flows (the
        # trigger is independent of the ``<tools>`` marker, so both
        # grounded and deflected turns hit it; SCAFFOLD_ALWAYS_TRIGGER
        # is checked first inside the classifier — the more specific
        # phrase wins, same convention as THINK_PARAS_TRIGGER).
        scaffold_flow = _scaffold_flow(body)
        if scaffold_flow is not None:
            # Request 1 (or EVERY request for the ALWAYS trigger): the
            # incident span as plain delta.content, 12-char chunks
            # (the span always spans ≥2 chunks — the filter's boundary
            # path), finish_reason "stop", no tool_calls, no reasoning.
            # Request 2 of the recovery trigger: the clean answer.
            answer = (
                SCAFFOLD_RECOVERY_ANSWER
                if scaffold_flow == "recovery"
                else SCAFFOLD_SPAN
            )
            return StreamingResponse(
                _sse_stream(answer, 0.0),
                media_type="text/event-stream",
                headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"},
            )
        # Phase 68 (search tool): the deterministic search marker flow —
        # checked BEFORE the phase-37 tool flow (the more specific
        # trigger phrase wins, same convention as THINK_PARAS_TRIGGER).
@@ -0,0 +1,438 @@
 """Phase 71 E2E (Playwright, mock-only): tool-scaffolding guardrails —
 the deterministic strip + one bounded recovery, through the real UI.
 Owner request (chat, 2026-09-03): the same incident as phase 70 — a
 deflected round streamed the model's raw
 ``<|tool_call_start|>[read(path='…')]<|tool_call_end|>`` chat-template
 markup into the UI although no tools were offered. The guardrail is
 DETERMINISTIC ONLY (no model in detection or repair): a streaming
 filter strips known scaffolding from ``delta.content`` server-side
 (``app/rag/scaffolding.py``), and a reply whose visible content ends
 up empty gets exactly ONE bounded recovery (``tools=None``, the
 harness correction folded into the system prompt); a second empty
 reply settles with the dedicated error frame.
 Run in isolation (DB must be up: ``podman compose up -d db``):
    uv run pytest tests/e2e/test_tool_scaffolding_guardrails.py -v --no-cov
 MOCK-ONLY suite: ``E2E_REAL_LLM=1`` is not supported — the gate is the
 deterministic scaffolding flows in ``tests/e2e/mock_llm.py`` (phase
 71), which are independent of the ``<tools>`` marker:
 * ``emit raw tool markup`` (``SCAFFOLD_TRIGGER``): request 1 streams
  ONLY the incident span as ``delta.content`` (split across the
  mock's 12-char chunks — the filter's boundary path), no structured
  ``tool_calls``, no reasoning; the follow-up request carrying
  ``CORRECTION_INSTRUCTION`` in the system prompt (the one bounded
  recovery) streams the clean ``SCAFFOLD_RECOVERY_ANSWER``.
 * ``always emit raw tool markup`` (``SCAFFOLD_ALWAYS_TRIGGER``): the
  scaffolding-only span on EVERY request (the recovery included) —
  the terminal malformed-reply path.
 Both triggers run on an EMPTY knowledge base: with zero chunks the
 honesty gate is LOW (cosine 0.0 < the E2E 0.30 threshold, no FTS
 hits), so every turn takes the DEFLECTED path — the exact path the
 2026-09-03 incident hit — where the filter + recovery live in
 ``app/api/chat.py``.
 Test → phase mapping (Playwright Mapping Rule):
 1. ``test_recovery_strips_scaffolding_and_streams_clean_answer`` — the
   recovery case: the turn settles (the composer re-enables, ``done``
   on the wire), the final answer bubble carries the recovery's clean
   text, ``document.body.innerText`` contains NEITHER
   ``tool_call_start`` nor ``tool_call_end`` (nor the raw
   ``[read(path=…`` fragment), and no error banner — and wire-level,
   no ``delta`` frame ever carries a scaffolding fragment (the strip
   happens server-side, not in the UI).
 2. ``test_terminal_scaffolding_lands_on_dedicated_error_app_stays_usable``
   — the terminal case: the existing error state renders with the
   dedicated copy ("The model returned a malformed reply — please try
   again."), no raw tokens in the DOM, no answer bubble, no ``done``
   and NO ``query_log`` row (the existing terminal-error semantics) —
   and the app stays usable: a follow-up plain question in the same
   session gets a normal deflected answer and the banner clears.
 3. ``test_plain_turn_never_recovers_and_streams_byte_clean`` — no
   false positive: a plain deflected question streams its
   first-request answer byte-clean (the concatenated delta text is
   EXACTLY the mock's deterministic deflection answer — not the
   recovery's), with no error state and no recovery request visible
   (the turn settles on the first request).
 """
 from __future__ import annotations
 import json
 import re
 import time
 from playwright.sync_api import Page, expect
 from sqlalchemy import select, text
 from app.db import SessionLocal
 from app.models import QueryLog
 from tests.e2e.mock_llm import SCAFFOLD_RECOVERY_ANSWER, compose_answer
 # --------------------------------------------------------------------------
 # Questions + the mock's deterministic expectations
 # --------------------------------------------------------------------------
 #: Carries ``SCAFFOLD_TRIGGER`` (and nothing else — the module-level
 #: asserts below pin the trigger exclusions).
 RECOVERY_Q = "Please emit raw tool markup in your reply to this."
 #: Carries ``SCAFFOLD_ALWAYS_TRIGGER`` — the scaffolding-only span on
 #: EVERY request, recovery included (the terminal path).
 TERMINAL_Q = "Please always emit raw tool markup in every reply."
 #: A plain question (the phase-67 deflection shape): no trigger at all.
 PLAIN_Q = "How do I bake sourdough bread?"
 for _q in (RECOVERY_Q, TERMINAL_Q, PLAIN_Q):
    for _other in (
        "read two documents",
        "search your documents",
        "write a long answer",
        "think in paragraphs",
        "think out loud",
        "show the end of your notes",
        "show me a table",
        "fail then answer",
        "always fail",
        "embed fail once",
        "pretend to think slowly",
        "use your tools",
    ):
        assert _other not in _q.lower(), f"{_other!r} unexpectedly in {_q!r}"
 assert "emit raw tool markup" in RECOVERY_Q.lower()
 assert "always emit raw tool markup" not in RECOVERY_Q.lower()
 assert "always emit raw tool markup" in TERMINAL_Q.lower()
 assert "emit raw tool markup" not in PLAIN_Q.lower()
 #: The terminal malformed-reply copy (``app.api.chat`` — the phase-71
 #: dedicated error frame; the mock's ALWAYS trigger is what makes the
 #: recovery come back empty).
 MALFORMED_ERROR_COPY = "The model returned a malformed reply — please try again."
 #: The mock's deterministic deflection phrase (the phase-67 pin).
 DEFLECT_PHRASE = r"haven't done anything like that"
 #: The raw scaffolding fragments that must NEVER reach the user — the
 #: span's tokens and the incident's argument fragment (``SCAFFOLD_SPAN``
 #: in the mock).
 RAW_TOKENS = ("tool_call_start", "tool_call_end", "[read(path=", "<|")
 def _expected_deflect_answer(question: str) -> str:
    """The mock's deterministic DEFLECT_MODE answer for *question*.
    Derived from the mock itself (``compose_answer`` on a synthetic
    body: the mode marker in the system prompt, the question as the
    user message, no tuning/KB sections — the suite truncates
    ``steering_notes`` / ``kb_overview``, so the real prompt carries
    none), so the byte-clean pin can never drift from the mock.
    """
    return compose_answer(
        {
            "messages": [
                {"role": "system", "content": "DEFLECT_MODE marker"},
                {"role": "user", "content": question},
            ]
        }
    )
 # --------------------------------------------------------------------------
 # DB reset (EMPTY knowledge base → every turn is deterministically
 # deflected: the LOW gate, the incident's path) + query_log reads
 # --------------------------------------------------------------------------
 def _reset_db_empty() -> None:
    """Truncate the KB (plus the prompt-shaping tables): an EMPTY
    knowledge base, so the honesty gate is LOW for every question
    (no chunks → cosine 0.0 < 0.30, fts_hits 0) — the deflected path
    where the phase-71 filter + recovery live, and the prompts stay
    byte-stable regardless of leftovers from other suites."""
    with SessionLocal() as db:
        db.execute(
            text("TRUNCATE chunks, documents, query_log, steering_notes, kb_overview")
        )
        db.commit()
 def _query_log_rows() -> list[QueryLog]:
    with SessionLocal() as db:
        return list(db.scalars(select(QueryLog)).all())
 # --------------------------------------------------------------------------
 # Page hooks (the SSE capture — the house pattern from
 # test_agent_document_tools.py / test_llm_retry.py)
 # --------------------------------------------------------------------------
 #: Captures the raw SSE ``data:`` payloads of the /api/chat stream
 #: (a response clone read in the background) — wire-level assertions
 #: independent of the UI rendering.
 SSE_HOOK = """
 () => {
  if (window.__sseInstalled) return;
  window.__sseInstalled = true;
  window.__sseFrames = [];
  const origFetch = window.fetch;
  window.fetch = async function (...args) {
    const res = await origFetch.apply(this, args);
    try {
      const url = typeof args[0] === 'string' ? args[0] : args[0].url;
      if (url.includes('/api/chat')) {
        res.clone().text().then((bodyText) => {
          for (const block of bodyText.split('\\n\\n')) {
            const line = block.trim();
            if (line.startsWith('data: ')) {
              window.__sseFrames.push(line.slice(6));
            }
          }
        });
      }
    } catch (e) { /* non-clonable responses: ignored */ }
    return res;
  };
 }
 """
 def _install_sse_hook(page: Page) -> None:
    page.evaluate(SSE_HOOK)
 def _drain_frames(page: Page, terminal: str = "done") -> list[dict]:
    """One turn's SSE frames: wait for that turn's *terminal* frame
    (``done`` — or ``error`` for the terminal case), then return EVERY
    frame captured since the last drain. The hook's background read
    appends the whole stream at once after it closes, so
    clearing-and-reading is race-free per turn: the terminal frame is
    observed only in the batch that carries the turn's full frame
    sequence."""
    deadline = time.monotonic() + 30.0
    while True:
        raw = page.evaluate(
            "() => { const f = window.__sseFrames || []; "
            "window.__sseFrames = []; return f; }"
        )
        parsed = [json.loads(line) for line in raw if line]
        if any(f.get("type") == terminal for f in parsed):
            return parsed
        if time.monotonic() > deadline:
            raise AssertionError(
                f"SSE hook captured no `{terminal}` frame (frames so far: "
                f"{len(parsed)}) — hook install failed?"
            )
        time.sleep(0.05)
 def _delta_text(frames: list[dict]) -> str:
    """The turn's answer text exactly as the wire carried it."""
    return "".join(f.get("text", "") for f in frames if f.get("type") == "delta")
 def _assert_no_raw_tokens_on_wire(frames: list[dict]) -> None:
    for frame in frames:
        if frame.get("type") != "delta":
            continue
        text = frame.get("text", "")
        for raw in RAW_TOKENS:
            assert raw not in text, f"raw scaffolding {raw!r} on the wire: {text!r}"
 def _submit(page: Page, question: str) -> None:
    page.fill("#message-input", question)
    page.click("#send-btn")
    # The user bubble lands synchronously with the submit handler.
    expect(page.locator(".msg.user .bubble").last).to_contain_text(question)
 def _wait_settled(page: Page) -> None:
    """The turn is complete: answer text in the bubble, button recovered.
    Phase 48: the label assertion carries the settle wait with an
    explicit timeout — the in-flight button is the enabled Stop control
    (never disabled), so ``to_be_enabled`` no longer blocks until the
    turn settles."""
    expect(page.locator(".msg.brain .bubble").last).not_to_have_text("", timeout=30_000)
    expect(page.locator("#send-btn")).to_be_enabled(timeout=30_000)
    expect(page.locator("#send-label")).to_have_text("Send", timeout=30_000)
 def _assert_no_raw_tokens(page: Page) -> None:
    """The user-visible contract: no raw scaffolding fragment anywhere
    in the rendered page (``document.body.innerText``)."""
    body = page.locator("body").inner_text()
    for raw in RAW_TOKENS:
        assert raw not in body, f"raw scaffolding {raw!r} leaked into the DOM"
 def _assert_no_error_banner(page: Page) -> None:
    """The turn settled through the normal done path — never the red
    role=alert error banner (the KB-offline banner is a separate,
    health-driven state the db_ready fixture keeps away)."""
    banner = page.locator("#kb-banner")
    expect(banner).to_be_hidden()
    expect(banner).not_to_have_attribute("role", "alert")
    expect(banner).not_to_have_class(re.compile(r"is-error"))
 # --------------------------------------------------------------------------
 # 1. The recovery case: the span is stripped server-side, the ONE
 #    bounded recovery answers, and no raw token ever reaches the DOM
 # --------------------------------------------------------------------------
 def test_recovery_strips_scaffolding_and_streams_clean_answer(
    page: Page, app_url: str, db_ready: None
 ) -> None:
    page.set_default_timeout(30_000)
    _reset_db_empty()
    page.goto(app_url)
    _install_sse_hook(page)
    _submit(page, RECOVERY_Q)
    _wait_settled(page)
    # The final answer bubble carries the RECOVERY's clean text — the
    # mock's deterministic recovery answer, proof the one bounded
    # recovery ran (request 2, with the correction in the system
    # prompt) and its answer is what the user saw.
    bubble = page.locator(".msg.brain .bubble").last
    expect(bubble).to_contain_text(SCAFFOLD_RECOVERY_ANSWER)
    # The deflected message state is intact — the turn took the LOW
    # path, the incident's path.
    expect(page.locator(".msg.brain").last).to_have_class(re.compile(r"is-deflected"))
    _assert_no_raw_tokens(page)
    _assert_no_error_banner(page)
    # Wire level: the strip happens SERVER-side — no `delta` frame ever
    # carries a scaffolding fragment, and the concatenated delta text is
    # EXACTLY the recovery answer (request 1's span produced zero delta
    # frames). No tool frames, no retry frames, no error frame; the
    # turn settled with `done` (deflected).
    frames = _drain_frames(page, terminal="done")
    assert [f for f in frames if f.get("type") == "delta"], (
        "the recovery answer must have streamed delta frames"
    )
    _assert_no_raw_tokens_on_wire(frames)
    assert _delta_text(frames) == SCAFFOLD_RECOVERY_ANSWER
    assert not [f for f in frames if f.get("type") == "tool"]
    assert not [f for f in frames if f.get("type") == "retry"]
    assert not [f for f in frames if f.get("type") == "error"]
    done = next(f for f in frames if f.get("type") == "done")
    assert done["deflected"] is True
    # The recovered turn settles durably: exactly one query_log row
    # (deflected — the incident's path).
    rows = _query_log_rows()
    assert len(rows) == 1, rows
    assert rows[0].question == RECOVERY_Q
    assert rows[0].deflected is True
 # --------------------------------------------------------------------------
 # 2. The terminal case: scaffolding twice → the dedicated error frame,
 #    no done, no query_log row — and the app stays usable
 # --------------------------------------------------------------------------
 def test_terminal_scaffolding_lands_on_dedicated_error_app_stays_usable(
    page: Page, app_url: str, db_ready: None
 ) -> None:
    page.set_default_timeout(30_000)
    _reset_db_empty()
    page.goto(app_url)
    _install_sse_hook(page)
    _submit(page, TERMINAL_Q)
    # BOTH requests (original + the one bounded recovery) came back
    # scaffolding-only: no clean content ever streamed, so the turn
    # settles with the EXISTING terminal error state — the banner
    # (role=alert) with the DEDICATED malformed-reply copy — and the
    # send button re-enabled (the banner path settles the state
    # machine, cf. the phase-67 exhaustion test).
    expect(page.locator("#kb-banner")).to_have_attribute(
        "role", "alert", timeout=60_000
    )
    expect(page.locator("#kb-banner")).to_contain_text(MALFORMED_ERROR_COPY)
    expect(page.locator("#send-btn")).to_be_enabled(timeout=30_000)
    expect(page.locator("#send-label")).to_have_text("Send", timeout=30_000)
    # No answer bubble was ever rendered (every streamed frame was
    # stripped server-side) and no raw token is anywhere in the DOM.
    expect(page.locator("#messages > .msg.brain")).to_have_count(0)
    _assert_no_raw_tokens(page)
    # Wire: the terminal error frame is LAST — no done, and NO delta
    # frame at all (both requests' content was pure scaffolding).
    frames = _drain_frames(page, terminal="error")
    assert frames[-1]["type"] == "error"
    assert MALFORMED_ERROR_COPY in frames[-1]["detail"]
    assert not [f for f in frames if f.get("type") == "done"]
    assert not [f for f in frames if f.get("type") == "delta"]
    # Terminal semantics (byte-for-byte the existing LLMError shape):
    # the turn writes no query_log row.
    assert _query_log_rows() == []
    # The app stays usable: a follow-up plain question (no trigger) in
    # the SAME session gets a normal streamed deflected answer, the
    # banner clears, and the wire is clean.
    _submit(page, PLAIN_Q)
    _wait_settled(page)
    bubble = page.locator(".msg.brain .bubble").last
    expect(bubble).to_contain_text(re.compile(DEFLECT_PHRASE, re.IGNORECASE))
    _assert_no_error_banner(page)
    _assert_no_raw_tokens(page)
    follow_frames = _drain_frames(page, terminal="done")
    assert next(f for f in follow_frames if f["type"] == "done")["deflected"] is True
    assert not [f for f in follow_frames if f.get("type") == "error"]
    _assert_no_raw_tokens_on_wire(follow_frames)
 # --------------------------------------------------------------------------
 # 3. No false positive: a plain turn streams its first-request answer
 #    byte-clean — no strip, no recovery request, no error state
 # --------------------------------------------------------------------------
 def test_plain_turn_never_recovers_and_streams_byte_clean(
    page: Page, app_url: str, db_ready: None
 ) -> None:
    page.set_default_timeout(30_000)
    _reset_db_empty()
    page.goto(app_url)
    _install_sse_hook(page)
    _submit(page, PLAIN_Q)
    _wait_settled(page)
    # Wire level: the plain deflected answer streams byte-clean from
    # the FIRST request — the concatenated delta text is EXACTLY the
    # mock's deterministic deflection answer (derived from the mock
    # itself above) and NOT the recovery answer (a recovery request
    # would have streamed that text instead).
    frames = _drain_frames(page, terminal="done")
    delta_text = _delta_text(frames)
    assert delta_text == _expected_deflect_answer(PLAIN_Q)
    assert SCAFFOLD_RECOVERY_ANSWER not in delta_text
    _assert_no_raw_tokens_on_wire(frames)
    # No recovery request is visible: no error frame, no retry frame
    # (the recovery is not a phase-67 retry), and the turn settled on
    # the first request with `done`.
    assert not [f for f in frames if f.get("type") == "error"]
    assert not [f for f in frames if f.get("type") == "retry"]
    done = next(f for f in frames if f.get("type") == "done")
    assert done["deflected"] is True
    # UI level: the deflected answer rendered, no error banner, no raw
    # token in the DOM.
    bubble = page.locator(".msg.brain .bubble").last
    expect(bubble).to_contain_text(re.compile(DEFLECT_PHRASE, re.IGNORECASE))
    _assert_no_error_banner(page)
    _assert_no_raw_tokens(page)
@@ -23,7 +23,7 @@ import asyncio
 import uuid
 from collections.abc import AsyncIterator, Iterator
 from copy import deepcopy
-from typing import Any, cast
+from typing import TYPE_CHECKING, Any, cast
 import pytest
 from sqlalchemy import delete, text
@@ -35,6 +35,9 @@ from app.rag import agent
 from app.rag.agent import AGENT_TOOLS, AgentHolder, run_agent
 from app.rag.llm import LLMClient, RetryPiece, StreamPiece, ToolCallPiece
 if TYPE_CHECKING:
    from app.rag.scaffolding import ScaffoldingFilter
 def _doc(db: Session, source: str, path: str, title: str, content: str) -> Document:
    doc = Document(
@@ -168,6 +171,7 @@ class ScriptedToolLLM:
        self,
        messages: list[dict[str, str]],
        tools: list[dict[str, Any]] | None = None,
        scaffolding: ScaffoldingFilter | None = None,  # phase 71 pass-through
    ) -> AsyncIterator[StreamPiece | ToolCallPiece]:
        self.requests.append((deepcopy(messages), deepcopy(tools)))
        if len(self.requests) == 1:
@@ -18,7 +18,7 @@ import math
 import re
 from collections.abc import Iterator
 from pathlib import Path
-from typing import Any
+from typing import TYPE_CHECKING, Any, cast
 import pytest
 from fastapi.testclient import TestClient
@@ -33,6 +33,9 @@ from app.rag.agent import AGENT_TOOLS
 from app.rag.importer import import_sources
 from app.rag.llm import EmbeddingError, LLMError, StreamPiece, ToolCallPiece
 if TYPE_CHECKING:
    from app.rag.scaffolding import ScaffoldingFilter
 FIXTURES = Path(__file__).resolve().parents[1] / "fixtures" / "docs"
 QUESTION = "How is my Kubernetes cluster set up?"
 OFF_TOPIC = "How do I bake sourdough bread?"
@@ -63,6 +66,7 @@ class FakeRagLLM:
        tool_script: list[list[StreamPiece | ToolCallPiece]] | None = None,
        embed_fail_count: int = 0,
        stream_fail_count: int = 0,
        answer_sequence: list[str] | None = None,
    ) -> None:
        self.settings = Settings(_env_file=None)  # pyright: ignore[reportCallIssue]
        self.embed_batches = 0
@@ -71,6 +75,11 @@ class FakeRagLLM:
        self.embed_error = embed_error
        self.stream_error = stream_error
        self.fail_mid_stream = fail_mid_stream
        #: Phase 71: per-request canned answers (the recovery matrix):
        #: request *i* (0-based, in ``seen_messages`` order) yields
        #: ``answer_sequence[i]``; once exhausted it falls back to
        #: ``answer``. ``None`` keeps the single-``answer`` behavior.
        self.answer_sequence = answer_sequence
        #: Phase 67: the first N ``embed_one`` calls raise an
        #: ``EmbeddingError`` (then succeed) — a dead-then-recovered
        #: embeddings endpoint for the retry loop.
@@ -117,17 +126,35 @@ class FakeRagLLM:
        self.question_embeds.append(text)
        return _token_vec(text)
    def _answer_for_request(self) -> str:
        """The canned answer for the request that was just recorded
        (phase 71 ``answer_sequence``; ``None`` → the single answer)."""
        if self.answer_sequence is None:
            return self.answer
        index = len(self.seen_messages) - 1
        if index < len(self.answer_sequence):
            return self.answer_sequence[index]
        return self.answer
    async def chat_stream(
        self,
        messages: list[dict[str, str]],
        tools: list[dict[str, Any]] | None = None,
        scaffolding: ScaffoldingFilter | None = None,
    ):
        """Typed stream (phase 17): ``thinking`` slices (same 12-char
        cadence as content) **before** the content pieces. With the
        default ``thinking=""`` this yields content-only pieces — today's
        behavior, new yield type. Phase 37: *tools* is the agent loop's
        ``tools=…`` passthrough (recorded in ``seen_tools``); a request
-        with tools consumes the next ``tool_script`` entry, if any."""
+        with tools consumes the next ``tool_script`` entry, if any.
        Phase 71: *scaffolding* mirrors ``LLMClient.chat_stream`` — the
        canned content pieces are fed through the caller's filter (an
        empty clean result yields nothing) and the held tail is flushed
        on normal completion, so a scaffolding-only canned answer streams
        zero content pieces and leaves ``stripped_chars`` behind for the
        recovery policy to key on. ``None`` (e.g. pre-phase callers) keeps
        the byte-identical raw path."""
        self.seen_messages.append(messages)
        self.seen_tools.append(tools)
        if self.stream_error is not None:
@@ -135,17 +162,46 @@ class FakeRagLLM:
        if self.stream_fail_count > 0:
            self.stream_fail_count -= 1
            raise LLMError("simulated pre-piece endpoint failure")
        mid_stream_drop = False
        raw: list[StreamPiece | ToolCallPiece]
        if tools is not None and self.tool_script:
-            for piece in self.tool_script.pop(0):
+            raw = self.tool_script.pop(0)
        elif self.fail_mid_stream:
            raw = [StreamPiece("content", "partial ")]
            mid_stream_drop = True
        else:
            answer = self._answer_for_request()
            raw = cast(
                "list[StreamPiece | ToolCallPiece]",
                [
                    StreamPiece("thinking", self.thinking[i : i + 12])
                    for i in range(0, len(self.thinking), 12)
                ]
                + [
                    StreamPiece("content", answer[i : i + 12])
                    for i in range(0, len(answer), 12)
                ],
            )
        if scaffolding is None:
            for piece in raw:
                yield piece
            if mid_stream_drop:
                raise LLMError("mid-stream dropout")
            return
-        if self.fail_mid_stream:
+        for piece in raw:
-            yield StreamPiece("content", "partial ")
+            if isinstance(piece, StreamPiece) and piece.kind == "content":
                cleaned = scaffolding.feed(piece.text)
                if cleaned:
                    yield StreamPiece("content", cleaned)
            else:
                yield piece
        if mid_stream_drop:
            # The tail is NOT flushed on a failed stream — the real
            # client only flushes a cleanly completed one.
            raise LLMError("mid-stream dropout")
-        for i in range(0, len(self.thinking), 12):
+        tail = scaffolding.flush()
-            yield StreamPiece("thinking", self.thinking[i : i + 12])
+        if tail:
-        for i in range(0, len(self.answer), 12):
+            yield StreamPiece("content", tail)
            yield StreamPiece("content", self.answer[i : i + 12])
@pytest.fixture()
@@ -621,6 +677,7 @@ def test_grounded_turn_streams_tool_frames_and_cites_read_doc(
    assert lines and "tool_calls=2" in lines[-1]
    assert "'docs/homelab/kubernetes.md'" in lines[-1]
    assert "'docs/homelab/backups.md'" in lines[-1]
    assert "scaffold_stripped=0" in lines[-1]  # phase 71: uniform clean-turn field
 def test_grounded_turn_streams_grep_tool_frames(
@@ -853,6 +910,7 @@ def test_zero_max_rounds_reproduce_pre_phase_single_request(
    assert "backups.md" not in row.sources
    lines = [r.getMessage() for r in caplog.records if "question=" in r.getMessage()]
    assert lines and "tool_calls=0" in lines[-1]
    assert "scaffold_stripped=0" in lines[-1]  # phase 71: uniform clean-turn field
 def test_tool_execution_db_failure_yields_error_event(
@@ -929,6 +987,7 @@ def test_embed_failure_retries_then_turn_completes(
    assert frames[-1]["type"] == "done"
    lines = [r.getMessage() for r in caplog.records if "question=" in r.getMessage()]
    assert lines and "retries=1" in lines[-1]
    assert "scaffold_stripped=0" in lines[-1]  # phase 71: uniform clean-turn field
 def test_embed_failure_exhausts_retries_then_terminal_error(
@@ -988,6 +1047,7 @@ def test_deflected_stream_retries_before_the_first_piece(
    assert flaky.seen_tools == [None, None]  # …byte-identical (no tools key)
    lines = [r.getMessage() for r in caplog.records if "question=" in r.getMessage()]
    assert lines and "retries=1" in lines[-1]
    assert "scaffold_stripped=0" in lines[-1]  # phase 71: uniform clean-turn field
 def test_deflected_stream_failure_after_first_frame_is_terminal(
@@ -1034,3 +1094,135 @@ def test_zero_retries_keep_the_pre_phase_wire_shape(
    assert frames[0]["type"] == "error"
    assert "embedding" in frames[0]["detail"]
    assert not any(f["type"] == "retry" for f in frames)
 # ---------- phase 71: the deterministic scaffolding guardrail (deflected path) ----------
 def _scaffold_span() -> str:
    """The raw span from the 2026-09-03 incident (the E2E mock's trigger,
    task 05) — a complete span the filter strips in full."""
    return "<|tool_call_start|>[read(path='/homelab/backup-notes.md')]<|tool_call_end|>"
 def test_deflected_scaffolding_only_reply_recovers_once(
    client,
    db,
    seeded_kb: FakeRagLLM,
    monkeypatch: pytest.MonkeyPatch,
    caplog: pytest.LogCaptureFixture,
 ) -> None:
    """(a) A deflected reply that is pure scaffolding streams ZERO delta
    frames (no raw tokens on the wire); the one bounded recovery —
    ``tools=None``, the correction folded into the single system prompt,
    a fresh filter, the same retry budget — streams the clean answer, the
    turn settles with ``done`` + a query_log row, and the log line counts
    the stripped chars (the recovery does not bump ``retries=N``)."""
    span = _scaffold_span()
    clean = "I don't have that on hand — try one of the chips below?"
    flaky = FakeRagLLM(answer_sequence=[span, clean])
    live = get_settings()
    monkeypatch.setattr(chat_api, "get_settings", lambda: _retry_settings(live))
    fastapi_app.dependency_overrides[chat_api.get_llm] = lambda: flaky
    try:
        caplog.set_level(logging.INFO, logger="app.chat")
        _, _, frames = _stream_chat(client, OFF_TOPIC)
    finally:
        fastapi_app.dependency_overrides.clear()
    # The raw tokens never reach the wire; the deltas reassemble to the
    # clean recovery answer.
    assert span not in json.dumps(frames)
    deltas = [f for f in frames if f["type"] == "delta"]
    assert "".join(d["text"] for d in deltas) == clean
    assert not any(f["type"] == "error" for f in frames)
    done = frames[-1]
    assert done["type"] == "done" and done["deflected"] is True
    # Exactly two requests: the stripped round + the one recovery, both
    # without a tools key…
    assert len(flaky.seen_messages) == 2
    assert flaky.seen_tools == [None, None]
    # …and the recovery's system prompt is the ORIGINAL deflected prompt
    # with the correction folded in (a single system message — the user
    # message stays last).
    recovered = flaky.seen_messages[1]
    assert len(recovered) == 2
    assert recovered[1] == {"role": "user", "content": OFF_TOPIC}
    first_system = flaky.seen_messages[0][0]["content"]
    assert "DEFLECT_MODE" in first_system
    assert recovered[0] == {
        "role": "system",
        "content": first_system + "\n" + agent.CORRECTION_INSTRUCTION,
    }
    # The turn settled normally: one query_log row…
    (row,) = db.scalars(select(QueryLog)).all()
    assert row.deflected is True
    # …and the log line carries the summed stripped count (the clean
    # recovery stripped nothing) with retries untouched.
    lines = [r.getMessage() for r in caplog.records if "question=" in r.getMessage()]
    assert lines and f"scaffold_stripped={len(span)}" in lines[-1]
    assert "retries=0" in lines[-1]  # the recovery is not an endpoint-retry
 def test_deflected_scaffolding_twice_settles_malformed(
    client, db, seeded_kb: FakeRagLLM, monkeypatch: pytest.MonkeyPatch
 ) -> None:
    """(b) The recovery answer is scaffolding again — a second empty
    reply is terminal: the DEDICATED error frame (the exact copy), no
    ``done``, no query_log row — byte-for-byte today's ``LLMError``
    terminal shape — and no third request (at most one recovery per
    turn)."""
    span = _scaffold_span()
    dead = FakeRagLLM(answer_sequence=[span, span])
    live = get_settings()
    monkeypatch.setattr(chat_api, "get_settings", lambda: _retry_settings(live))
    fastapi_app.dependency_overrides[chat_api.get_llm] = lambda: dead
    try:
        _, _, frames = _stream_chat(client, OFF_TOPIC)
    finally:
        fastapi_app.dependency_overrides.clear()
    assert [f["type"] for f in frames] == ["error"]
    assert frames[0]["detail"] == (
        "The model returned a malformed reply — please try again."
    )
    assert set(frames[0].keys()) == {"type", "detail"}  # the contract shape
    assert span not in json.dumps(frames)
    assert not any(f["type"] == "done" for f in frames)
    assert db.scalars(select(QueryLog)).all() == []
    assert len(dead.seen_messages) == 2  # round + one recovery — no more
    assert dead.seen_tools == [None, None]
 def test_deflected_mixed_scaffolding_and_content_needs_no_recovery(
    client,
    db,
    seeded_kb: FakeRagLLM,
    monkeypatch: pytest.MonkeyPatch,
    caplog: pytest.LogCaptureFixture,
 ) -> None:
    """(c) Real visible content plus scaffolding: the clean remainder
    streams (no raw tokens on the wire), NO recovery runs, and the log
    line counts the stripped span (``scaffold_stripped>0``)."""
    span = _scaffold_span()
    mixed = f"I don't have that. {span} Try the chips below?"
    flaky = FakeRagLLM(answer=mixed)
    live = get_settings()
    monkeypatch.setattr(chat_api, "get_settings", lambda: _retry_settings(live))
    fastapi_app.dependency_overrides[chat_api.get_llm] = lambda: flaky
    try:
        caplog.set_level(logging.INFO, logger="app.chat")
        _, _, frames = _stream_chat(client, OFF_TOPIC)
    finally:
        fastapi_app.dependency_overrides.clear()
    assert span not in json.dumps(frames)
    deltas = [f for f in frames if f["type"] == "delta"]
    assert "".join(d["text"] for d in deltas) == "I don't have that.  Try the chips below?"
    assert not any(f["type"] == "error" for f in frames)
    assert frames[-1]["type"] == "done"
    assert len(flaky.seen_messages) == 1  # the clean content stands — no recovery
    lines = [r.getMessage() for r in caplog.records if "question=" in r.getMessage()]
    assert lines and f"scaffold_stripped={len(span)}" in lines[-1]
@@ -32,7 +32,7 @@ import logging
 import uuid
 from collections.abc import AsyncGenerator, AsyncIterator
 from copy import deepcopy
-from typing import Any, cast
+from typing import TYPE_CHECKING, Any, cast
 import pytest
 from sqlalchemy.orm import Session
@@ -43,11 +43,15 @@ from app.rag import agent
 from app.rag.agent import (
    AGENT_TOOLS,
    AgentHolder,
    MalformedReplyError,
    run_agent,
 )
 from app.rag.llm import LLMClient, LLMError, RetryPiece, StreamPiece, ToolCallPiece
 from app.rag.prompts import TOOLS_SECTION, _base, build_deflect_prompt, build_high_prompt
 if TYPE_CHECKING:
    from app.rag.scaffolding import ScaffoldingFilter
 def _settings(**kwargs: Any) -> Settings:
    kwargs.setdefault("_env_file", None)
@@ -68,7 +72,13 @@ def _doc(source: str, path: str, title: str = "Title", content: str = "CONTENT")
 class ScriptedLLM:
    """Canned stream sequences; records every ``chat_stream`` request so
-    the tests can assert on the messages and the ``tools`` passthrough."""
+    the tests can assert on the messages and the ``tools`` passthrough.
    Phase 71: when the caller passes a ``ScaffoldingFilter``, the canned
    content pieces are fed through it exactly like
    ``LLMClient.chat_stream`` (an empty clean result yields nothing; the
    held tail is flushed on normal completion) — so a scaffolding-only
    canned round streams no content pieces and leaves ``stripped_chars``
    behind for the recovery policy to key on."""
    def __init__(self, *streams: list[StreamPiece | ToolCallPiece]) -> None:
        self.streams: list[list[StreamPiece | ToolCallPiece]] = list(streams)
@@ -78,12 +88,26 @@ class ScriptedLLM:
        self,
        messages: list[dict[str, str]],
        tools: list[dict[str, Any]] | None = None,
        scaffolding: ScaffoldingFilter | None = None,
    ) -> AsyncIterator[StreamPiece | ToolCallPiece]:
        self.requests.append((deepcopy(messages), tools))
        if not self.streams:
            raise AssertionError("ScriptedLLM ran out of canned streams")
-        for piece in self.streams.pop(0):
+        pieces = self.streams.pop(0)
-            yield piece
+        if scaffolding is None:
            for piece in pieces:
                yield piece
            return
        for piece in pieces:
            if isinstance(piece, StreamPiece) and piece.kind == "content":
                cleaned = scaffolding.feed(piece.text)
                if cleaned:
                    yield StreamPiece("content", cleaned)
            else:
                yield piece
        tail = scaffolding.flush()
        if tail:
            yield StreamPiece("content", tail)
 async def _run(
@@ -1102,6 +1126,7 @@ class FailingLLM:
        self,
        messages: list[dict[str, str]],
        tools: list[dict[str, Any]] | None = None,
        scaffolding: ScaffoldingFilter | None = None,  # phase 71: fed like the real client
    ) -> AsyncIterator[StreamPiece | ToolCallPiece]:
        index = len(self.requests)
        pieces, error = (
@@ -1112,19 +1137,35 @@ class FailingLLM:
        self.requests.append(
            (deepcopy(messages), deepcopy(tools) if tools is not None else None)
        )
-        return self._attempt(index, pieces, error)
+        return self._attempt(index, pieces, error, scaffolding)
    async def _attempt(
        self,
        index: int,
        pieces: list[StreamPiece | ToolCallPiece],
        error: Exception | None,
        scaffolding: ScaffoldingFilter | None,
    ) -> AsyncIterator[StreamPiece | ToolCallPiece]:
        try:
            for piece in pieces:
-                yield piece
+                if (
                    scaffolding is not None
                    and isinstance(piece, StreamPiece)
                    and piece.kind == "content"
                ):
                    cleaned = scaffolding.feed(piece.text)
                    if cleaned:
                        yield StreamPiece("content", cleaned)
                else:
                    yield piece
            if error is not None:
                # The tail is NOT flushed on a failed attempt — the real
                # client only flushes a cleanly completed stream.
                raise error
            if scaffolding is not None:
                tail = scaffolding.flush()
                if tail:
                    yield StreamPiece("content", tail)
        finally:
            self.closed.append(index)
@@ -1402,11 +1443,16 @@ def test_high_prompt_tools_section_with_notes_and_kb() -> None:
 def test_low_prompt_is_byte_identical_and_tool_free() -> None:
    # Phase 71: the LOW prompt carries the owner-permitted plain-text
    # line after the DEFLECT_MODE sentence (the marker-keying contract
    # is unchanged; the line must not leak into the HIGH prompt —
    # pinned in tests/unit/test_prompts.py).
    expected = (
        _base("LOW")
        + "\nDEFLECT_MODE: retrieval was weak — the titles below are the closest "
        "your notes come to the question. They are titles only; do not pretend "
        "they answer it. Use them to propose 2-3 alternative questions.\n"
        "Reply in plain text only — you have no tools in this mode.\n"
        + "- T1\n- T2"
    )
    assert build_deflect_prompt(["T1", "T2"]) == expected
@@ -1418,3 +1464,253 @@ def test_low_prompt_is_byte_identical_and_tool_free() -> None:
    ):
        assert "<tools>" not in prompt
        assert TOOLS_SECTION not in prompt
 # ---------- phase 71: the scaffolding recovery policy (deterministic only) ----------
 #: The raw span from the 2026-09-03 incident (the E2E mock's trigger,
 #: task 05) — a complete span the filter strips in full.
 _INCIDENT_SPAN = (
    "<|tool_call_start|>[read(path='/homelab/backup-notes.md')]<|tool_call_end|>"
 )
 def test_correction_instruction_is_the_harness_constant() -> None:
    """Verbatim constant: the E2E mock (task 05) keys on a stable
    substring of it, so it must not drift."""
    assert agent.CORRECTION_INSTRUCTION == (
        "Your previous reply contained raw tool-call markup, which is not "
        "interpreted here. Answer the user's question directly in plain "
        "text — no tool syntax."
    )
 def test_malformed_reply_error_subclasses_llm_error() -> None:
    assert issubclass(MalformedReplyError, LLMError)
    assert not issubclass(LLMError, MalformedReplyError)
 def test_scaffolding_only_round_gets_exactly_one_recovery(
    monkeypatch: pytest.MonkeyPatch,
    caplog: pytest.LogCaptureFixture,
 ) -> None:
    """A round whose visible content is pure scaffolding → exactly TWO
    model requests: the normal round, then the ONE recovery —
    ``tools=None`` with :data:`CORRECTION_INSTRUCTION` folded into the
    original single system message (the user message stays last). The
    clean recovery answer ends the turn, the holder is untouched by the
    recovery, and the strip was captured in the warning log."""
    holder = AgentHolder()
    llm = ScriptedLLM(
        [StreamPiece("content", _INCIDENT_SPAN)],
        [StreamPiece("content", "The clean recovery answer.")],
    )
    with caplog.at_level(logging.WARNING, logger="app.agent"):
        pieces = asyncio.run(_run(llm, holder, _settings()))
    # Nothing of the round's scaffolding was yielded — only the recovery.
    assert pieces == [StreamPiece("content", "The clean recovery answer.")]
    assert len(llm.requests) == 2  # the round + the one recovery
    # The round: the original system prompt, the tools offered.
    assert llm.requests[0][1] == AGENT_TOOLS
    assert llm.requests[0][0] == [
        {"role": "system", "content": "SYSTEM_PROMPT"},
        {"role": "user", "content": "QUESTION"},
    ]
    # The recovery: no tools, a SINGLE system message carrying the folded
    # correction, the user message last.
    assert llm.requests[1][1] is None
    assert llm.requests[1][0] == [
        {
            "role": "system",
            "content": "SYSTEM_PROMPT\n" + agent.CORRECTION_INSTRUCTION,
        },
        {"role": "user", "content": "QUESTION"},
    ]
    # The recovery is a fixed policy, not a conversation: the holder is
    # untouched by it.
    assert holder.read_docs == [] and holder.tool_calls == 0
    # The turn total feeds the API layer's ``scaffold_stripped=N`` field.
    assert holder.scaffold_stripped == len(_INCIDENT_SPAN)
    # One strip warning per stripped span, the span truncated to 200 chars.
    strip_logs = [
        r
        for r in caplog.records
        if r.levelno == logging.WARNING and r.getMessage().startswith("agent: stripped")
    ]
    assert len(strip_logs) == 1
    message = strip_logs[0].getMessage()
    assert message.startswith(
        f"agent: stripped {len(_INCIDENT_SPAN)} chars of tool-scaffolding in round 1:"
    )
    assert _INCIDENT_SPAN[:200] in message
 def test_scaffolding_twice_settles_with_malformed_reply(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    """The recovery answer is scaffolding again (a second empty reply) —
    terminal: :class:`MalformedReplyError` (an :class:`LLMError` subclass)
    after exactly two requests — no third request, no recovery of a
    recovery."""
    holder = AgentHolder()
    llm = ScriptedLLM(
        [StreamPiece("content", _INCIDENT_SPAN)],
        [StreamPiece("content", _INCIDENT_SPAN)],
    )
    with pytest.raises(MalformedReplyError) as excinfo:
        asyncio.run(_run(llm, holder, _settings()))
    assert isinstance(excinfo.value, LLMError)
    assert len(llm.requests) == 2  # exactly one recovery per turn
    assert llm.requests[1][1] is None
    assert agent.CORRECTION_INSTRUCTION in llm.requests[1][0][0]["content"]
    assert holder.scaffold_stripped == 2 * len(_INCIDENT_SPAN)
 def test_scaffolding_with_real_content_needs_no_recovery(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    """A round with real visible content PLUS scaffolding: the clean
    content stands — one request only, the clean remainder yielded (no
    raw tokens), no correction in any system prompt."""
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            StreamPiece("content", "Here it is: "),
            StreamPiece("content", _INCIDENT_SPAN),
            StreamPiece("content", " hope that helps."),
        ],
    )
    pieces = asyncio.run(_run(llm, holder, _settings()))
    assert [p for p in pieces if isinstance(p, StreamPiece)] == [
        StreamPiece("content", "Here it is: "),
        StreamPiece("content", " hope that helps."),
    ]
    assert len(llm.requests) == 1  # no recovery
    for messages, _tools in llm.requests:
        assert all(agent.CORRECTION_INSTRUCTION not in m["content"] for m in messages)
    assert holder.scaffold_stripped == len(_INCIDENT_SPAN)
 def test_clean_turn_carries_no_correction_and_no_strip(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    """A clean turn: one request, no correction in any system prompt,
    zero stripped (the log field stays 0 — uniform)."""
    holder = AgentHolder()
    llm = ScriptedLLM(
        [StreamPiece("thinking", "hmm "), StreamPiece("content", "a clean answer")]
    )
    pieces = asyncio.run(_run(llm, holder, _settings()))
    assert pieces == [
        StreamPiece("thinking", "hmm "),
        StreamPiece("content", "a clean answer"),
    ]
    assert len(llm.requests) == 1
    for messages, _tools in llm.requests:
        assert all(agent.CORRECTION_INSTRUCTION not in m["content"] for m in messages)
    assert holder.scaffold_stripped == 0
 def test_empty_round_without_a_strip_keeps_today_behavior(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    """Round content 0 with NOTHING stripped (an empty/thinking-only
    answer) → return as today — no recovery, no error."""
    holder = AgentHolder()
    llm = ScriptedLLM([StreamPiece("thinking", "nothing to say, honestly")])
    pieces = asyncio.run(_run(llm, holder, _settings()))
    assert pieces == [StreamPiece("thinking", "nothing to say, honestly")]
    assert len(llm.requests) == 1
    assert holder.scaffold_stripped == 0
 def test_scaffolding_round_with_tool_calls_needs_no_recovery(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    """A scaffolding-only round that ALSO carried tool calls: the tool
    ran, and the policy keys on the no-calls exit only — no recovery (the
    next round is a normal tools-offered round carrying the tool
    history)."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            StreamPiece("content", _INCIDENT_SPAN),
            ToolCallPiece(id="call_1", name="ls", arguments={}),
        ],
        [StreamPiece("content", "the answer after the tool")],
    )
    pieces = asyncio.run(_run(llm, holder, _settings()))
    # The round's scaffolding was stripped (no raw delta), the tool frame
    # and the next round's answer flowed on.
    assert pieces == [
        ToolCallPiece(id="call_1", name="ls", arguments={}),
        StreamPiece("content", "the answer after the tool"),
    ]
    assert len(llm.requests) == 2
    assert llm.requests[1][1] == AGENT_TOOLS  # a normal round, not a recovery
    for messages, _tools in llm.requests:
        assert all(
            m["content"] is None or agent.CORRECTION_INSTRUCTION not in m["content"]
            for m in messages
        )
    assert holder.tool_calls == 1
    assert holder.scaffold_stripped == len(_INCIDENT_SPAN)
 def test_recovery_after_tool_rounds_keeps_the_history(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    """A scaffolding-only answer round after a tool round: the recovery
    keeps the SINGLE (folded) system message at the front and the tool
    history intact behind it — no second system message, no duplicated
    correction."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={})],
        [StreamPiece("content", _INCIDENT_SPAN)],
        [StreamPiece("content", "recovered after a tool round")],
    )
    pieces = asyncio.run(_run(llm, holder, _settings()))
    assert pieces == [
        ToolCallPiece(id="call_1", name="ls", arguments={}),
        StreamPiece("content", "recovered after a tool round"),
    ]
    assert len(llm.requests) == 3  # tool round + stripped round + recovery
    assert llm.requests[2][1] is None
    recovered = llm.requests[2][0]
    assert recovered[0] == {
        "role": "system",
        "content": "SYSTEM_PROMPT\n" + agent.CORRECTION_INSTRUCTION,
    }
    assert recovered[1] == {"role": "user", "content": "QUESTION"}
    assert len(recovered) == 4
    assert recovered[2]["role"] == "assistant"
    assert recovered[3] == {
        "role": "tool",
        "tool_call_id": "call_1",
        "content": "1 documents:\nsource: S | path: a.md | title: A",
    }
    assert sum(1 for m in recovered if m["role"] == "system") == 1
    assert holder.scaffold_stripped == len(_INCIDENT_SPAN)
 def test_forced_final_scaffolding_only_settles_malformed(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    """The round-cap forced final (``tools=None``) is filtered too: a
    scaffolding-only forced answer never reaches the user raw — the turn
    settles with :class:`MalformedReplyError` (the same terminal
    semantics; this turn used no recovery, so nothing is doubled up)."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={})],
        [StreamPiece("content", _INCIDENT_SPAN)],
    )
    with pytest.raises(MalformedReplyError):
        asyncio.run(_run(llm, holder, _settings(agent_max_rounds=1)))
    assert len(llm.requests) == 2
    assert llm.requests[1][1] is None  # the forced final — no recovery after it
    assert holder.scaffold_stripped == len(_INCIDENT_SPAN)
@@ -11,7 +11,7 @@ from __future__ import annotations
 import json
 import uuid
 from collections.abc import Iterator
-from typing import Any
+from typing import TYPE_CHECKING, Any
 import pytest
 from fastapi.testclient import TestClient
@@ -25,6 +25,9 @@ from app.rag.llm import StreamPiece
 from app.rag.retriever import RetrievedChunk, weak_hit_titles
 from app.rag.suggestions import MAX_SUGGESTIONS, derive_suggestions
 if TYPE_CHECKING:
    from app.rag.scaffolding import ScaffoldingFilter
 ANSWER = "I haven't done anything like that — try one of these instead!"
 #: A small KB outline standing in for the lite-generated one (phase 31).
@@ -332,6 +335,11 @@ def test_low_prompt_has_titles_only_no_content() -> None:
    assert "<relevance>LOW</relevance>" in prompt
    assert "DEFLECT_MODE" in prompt
    assert "HONESTY GATE" in prompt  # the LOW rule is what the model follows
    # Phase 71 (owner-permitted 2026-09-03): the deflection plain-text
    # line — the LOW turn offers no tools, so any tool markup there is
    # always wrong (prevention at the prompt; the filter + recovery is
    # the backstop).
    assert "Reply in plain text only — you have no tools in this mode." in prompt
    assert "- Kubernetes Homelab Cluster" in prompt
    assert "- Backup Strategy" in prompt
    assert "ALPHA_DOC_CONTENT" not in prompt
@@ -347,6 +355,8 @@ def test_high_path_unaffected() -> None:
    assert plan.suggestions == []
    assert "<relevance>HIGH</relevance>" in plan.system_prompt
    assert "DEFLECT_MODE" not in plan.system_prompt
    # Phase 71: the deflection plain-text line never leaks into HIGH.
    assert "Reply in plain text only" not in plan.system_prompt
    assert "ALPHA_DOC_CONTENT" in plan.system_prompt
    assert "BETA_DOC_CONTENT" in plan.system_prompt
    assert [d.title for d in plan.docs] == ["Kubernetes Homelab Cluster", "Backup Strategy"]
@@ -438,6 +448,7 @@ class _CannedLLM:
        self,
        messages: list[dict[str, str]],
        tools: list[dict[str, Any]] | None = None,
        scaffolding: ScaffoldingFilter | None = None,  # phase 71 pass-through
    ):
        self.seen.append(messages)
        self.seen_tools.append(tools)
@@ -27,6 +27,7 @@ from app.rag.llm import (
    ToolCallPiece,
    chat_stream_retried,
 )
 from app.rag.scaffolding import ScaffoldingFilter
 def _settings(**kwargs: Any) -> Settings:
@@ -337,13 +338,19 @@ class _FakeCompletions:
        self.completion = completion
        self.kwargs: dict | None = None
        self.chat_kwargs: dict | None = None
        #: Every SDK-shaped stream handed out — teardown tests assert the
        #: phase-48 ``close()`` on them (phase 71 task 02: with/without
        #: a filter, the teardown path is the same object).
        self.streams: list[_FakeChatStream] = []
    async def create(self, **kwargs) -> _FakeChatStream | _FakeCompletion:
        self.kwargs = kwargs
        if self.fail is not None:
            raise self.fail
        if kwargs.get("stream"):
-            return _FakeChatStream(self.chunks)
+            stream = _FakeChatStream(self.chunks)
            self.streams.append(stream)
            return stream
        self.chat_kwargs = kwargs
        assert self.completion is not None
        return self.completion
@@ -721,6 +728,146 @@ def test_chat_stream_non_object_tool_arguments_raise_llm_error() -> None:
        asyncio.run(drain())
 # ---------- scaffolding filter integration (phase 71, task 02) ----------
 #: The incident's raw span (2026-09-03) — the filter's reason to exist.
 _INCIDENT_SPAN = "<|tool_call_start|>[read(path='/homelab/backup-notes.md')]<|tool_call_end|>"
 def _collect_filtered(
    llm: LLMClient,
    messages: list[dict[str, str]],
    scaffolding: ScaffoldingFilter | None,
 ) -> list[StreamPiece]:
    """Collect pieces from a tools-less filtered stream (phase 71): without
    tools, no ToolCallPiece can appear (the phase-37 contract)."""
    async def run() -> list[StreamPiece]:
        pieces = [p async for p in llm.chat_stream(messages, scaffolding=scaffolding)]
        assert all(isinstance(p, StreamPiece) for p in pieces)
        return cast("list[StreamPiece]", pieces)
    return asyncio.run(run())
 def test_chat_stream_with_filter_strips_span_mid_stream() -> None:
    """A span mid-stream never reaches the pieces: the clean remainder
    flows, ``stripped_chars`` is exact, and no piece carries scaffolding."""
    f = ScaffoldingFilter()
    llm, _ = _make_stream_client(
        [_chunk("Hello "), _chunk(_INCIDENT_SPAN + " world"), _chunk("!")]
    )
    pieces = _collect_filtered(llm, _RETRY_MSGS, f)
    assert [(p.kind, p.text) for p in pieces] == [
        ("content", "Hello "),
        ("content", " world"),
        ("content", "!"),
    ]
    assert f.stripped_chars == len(_INCIDENT_SPAN)
    assert all("<|" not in p.text for p in pieces if p.kind == "content")
 def test_chat_stream_with_filter_span_split_across_chunks() -> None:
    """A span split across two chunks never emits a partial marker —
    nothing leaks until the span completes, then the clean text on both
    sides flows and the whole span counts as stripped."""
    span = _INCIDENT_SPAN
    cut = len("<|tool_call_start|>")  # split right after the start token
    f = ScaffoldingFilter()
    llm, _ = _make_stream_client([_chunk("A " + span[:cut]), _chunk(span[cut:] + " B")])
    pieces = _collect_filtered(llm, _RETRY_MSGS, f)
    assert [(p.kind, p.text) for p in pieces] == [
        ("content", "A "),
        ("content", " B"),
    ]
    assert f.stripped_chars == len(span)
 def test_chat_stream_with_filter_yields_nothing_for_pure_scaffolding() -> None:
    """A content stream of pure scaffolding yields ZERO content pieces —
    an empty clean result yields nothing (no empty ``delta`` frames)."""
    f = ScaffoldingFilter()
    llm, _ = _make_stream_client(
        [_chunk(_INCIDENT_SPAN), _chunk("<|tool_calls|>")]
    )
    pieces = _collect_filtered(llm, _RETRY_MSGS, f)
    assert pieces == []
    assert f.stripped_chars == len(_INCIDENT_SPAN) + len("<|tool_calls|>")
 def test_chat_stream_with_filter_leaves_thinking_raw() -> None:
    """Locked (phase 71): the scratchpad stays raw — a span inside
    ``reasoning_content`` is yielded verbatim and never counts as
    stripped."""
    f = ScaffoldingFilter()
    llm, _ = _make_stream_client(
        [_chunk("", reasoning=_INCIDENT_SPAN), _chunk("ok")]
    )
    pieces = _collect_filtered(llm, _RETRY_MSGS, f)
    assert [(p.kind, p.text) for p in pieces] == [
        ("thinking", _INCIDENT_SPAN),
        ("content", "ok"),
    ]
    assert f.stripped_chars == 0
 def test_chat_stream_without_filter_is_the_raw_path() -> None:
    """``scaffolding=None`` (the default, and the explicit opt-out): a span
    in content is yielded verbatim — byte-identical to the pre-phase-71
    raw path for callers that do not pass a filter."""
    llm, _ = _make_stream_client([_chunk(_INCIDENT_SPAN)])
    explicit_none = _collect_filtered(llm, _RETRY_MSGS, None)
    default = asyncio.run(_collect(llm, _RETRY_MSGS))
    assert explicit_none == default == [StreamPiece("content", _INCIDENT_SPAN)]
 def test_chat_stream_flushed_tail_precedes_tool_call_pieces() -> None:
    """Content-before-tools wire convention (phase 71): a stream that ends
    with a held filter tail (a partial marker at EOF — flushed as-is) +
    tool_calls deltas yields the flushed tail content piece BEFORE the
    materialized ``ToolCallPiece``."""
    f = ScaffoldingFilter()
    llm, _ = _make_stream_client(
        [
            _chunk("done <|tool_call_st"),  # held: a live prefix of the start token
            _chunk(
                None,
                tool_calls=[_tool_call(0, id="call_t", name="ls")],
                finish_reason="tool_calls",
            ),
        ]
    )
    async def run() -> list[StreamPiece | ToolCallPiece]:
        return [
            p
            async for p in llm.chat_stream(_RETRY_MSGS, tools=_AGENT_TOOLS, scaffolding=f)
        ]
    pieces = asyncio.run(run())
    assert pieces == [
        StreamPiece("content", "done "),
        StreamPiece("content", "<|tool_call_st"),
        ToolCallPiece(id="call_t", name="ls", arguments={}),
    ]
 def test_chat_stream_abandon_with_filter_closes_stream() -> None:
    """Phase-48 teardown is independent of the phase-71 filter: a consumer
    abandon mid-filter still closes the endpoint stream exactly once."""
    f = ScaffoldingFilter()
    llm, completions = _make_stream_client([_chunk(f"piece {i} ") for i in range(1, 6)])
    async def run() -> None:
        gen = llm.chat_stream(_RETRY_MSGS, scaffolding=f)
        first = await gen.__anext__()
        assert isinstance(first, StreamPiece)
        assert first.text == "piece 1 "
        await gen.aclose()  # the consumer stops after the first piece
    asyncio.run(run())
    assert completions.streams[0].close_calls == 1
 # ---------- one-shot chat: LLMClient.chat (phase 30, task 01) ----------
@@ -826,11 +973,15 @@ class _ScriptedClient(LLMClient):
        ] = []
        #: Indices of attempts whose stream teardown has run.
        self.closed: list[int] = []
        #: The caller-owned filter each attempt's ``chat_stream`` received
        #: (phase 71 task 02 — the retry primitive forwards it).
        self.scaffoldings: list[ScaffoldingFilter | None] = []
    def chat_stream(
        self,
        messages: list[dict[str, str]],
        tools: list[dict[str, Any]] | None = None,
        scaffolding: ScaffoldingFilter | None = None,
    ) -> AsyncGenerator[StreamPiece | ToolCallPiece, None]:
        index = len(self.request_args)
        pieces, error = (
@@ -841,15 +992,30 @@ class _ScriptedClient(LLMClient):
        self.request_args.append(
            (list(messages), list(tools) if tools is not None else None)
        )
-        return self._attempt(index, pieces, error)
+        self.scaffoldings.append(scaffolding)
        return self._attempt(index, pieces, error, scaffolding)
    async def _attempt(
-        self, index: int, pieces: list[StreamPiece | ToolCallPiece],
+        self,
        index: int,
        pieces: list[StreamPiece | ToolCallPiece],
        error: Exception | None,
        scaffolding: ScaffoldingFilter | None,
    ) -> AsyncGenerator[StreamPiece | ToolCallPiece, None]:
        try:
            for piece in pieces:
-                yield piece
+                if (
                    scaffolding is not None
                    and isinstance(piece, StreamPiece)
                    and piece.kind == "content"
                ):
                    # Emulate the real contract (phase 71): content feeds
                    # the filter, an empty clean result yields nothing.
                    cleaned = scaffolding.feed(piece.text)
                    if cleaned:
                        yield StreamPiece("content", cleaned)
                else:
                    yield piece
            if error is not None:
                raise error
        finally:
@@ -875,12 +1041,18 @@ def _collect_retried(
    tools: list[dict[str, Any]] | None = None,
    retries: int,
    delay: float,
    scaffolding: ScaffoldingFilter | None = None,
 ) -> list[StreamPiece | ToolCallPiece | RetryPiece]:
    async def run() -> list[StreamPiece | ToolCallPiece | RetryPiece]:
        return [
            p
            async for p in chat_stream_retried(
-                client, messages, tools=tools, retries=retries, delay=delay
+                client,
                messages,
                tools=tools,
                retries=retries,
                delay=delay,
                scaffolding=scaffolding,
            )
        ]
@@ -1016,6 +1188,40 @@ def test_retried_zero_delay_still_notifies(monkeypatch: pytest.MonkeyPatch) -> N
    assert sleeps == [0]
 def test_retried_forwards_the_filter_to_every_attempt(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    """Phase 71: the caller-owned filter reaches EVERY attempt's
    ``chat_stream`` — the same object on the dead attempt and the
    surviving one."""
    client = _ScriptedClient(
        [([], LLMError("down")), ([StreamPiece("content", "ok")], None)]
    )
    f = ScaffoldingFilter()
    _record_sleeps(monkeypatch)
    pieces = _collect_retried(client, _RETRY_MSGS, retries=1, delay=0, scaffolding=f)
    assert pieces == [RetryPiece(2, 2), StreamPiece("content", "ok")]
    assert client.scaffoldings == [f, f]
 def test_retried_reuses_the_unfed_filter_after_a_dead_attempt(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    """Phase 71: reusing the SAME filter across attempts is safe by
    construction — the dead attempt emitted no piece, so the filter was
    never fed; the surviving attempt filters through it as if fresh."""
    span = _INCIDENT_SPAN
    client = _ScriptedClient(
        [([], LLMError("down")), ([StreamPiece("content", span + " clean")], None)]
    )
    f = ScaffoldingFilter()
    _record_sleeps(monkeypatch)
    pieces = _collect_retried(client, _RETRY_MSGS, retries=1, delay=0, scaffolding=f)
    assert pieces == [RetryPiece(2, 2), StreamPiece("content", " clean")]
    assert f.stripped_chars == len(span)
    assert client.scaffoldings == [f, f]
 def test_retried_abandon_mid_attempt_closes_the_attempt_stream() -> None:
    """Consumer abandon at a mid-attempt piece (the stop-generation path,
    phase 48): no exception leaks and the attempt's stream is torn down
@@ -5,6 +5,11 @@ lite-generated KB outline): its own builder contract (empty → ``""``,
 char budget + ``[…truncated…]`` marker, pathological budgets), its
 placement between ``<relevance>`` and ``<tuning>`` in both modes, and
 the byte-identical-when-absent convention (phase 15 precedent).
 And the phase-71 deflection plain-text line (owner-permitted
 2026-09-03): the LOW prompt = pre-phase text + exactly the one new
 line; the ``DEFLECT_MODE`` marker-keying contract is unchanged and
 the line never leaks into the HIGH prompt.
 """
 from __future__ import annotations
@@ -129,11 +134,15 @@ def test_zero_note_prompt_is_byte_identical_to_pre_steering() -> None:
    assert build_high_prompt([doc]) == (
        _base("HIGH") + "\n<documents>\n" + block + "\n</documents>" + "\n" + TOOLS_SECTION
    )
    # Phase 71: the LOW prompt carries the owner-permitted plain-text
    # line after the DEFLECT_MODE sentence (the marker-keying contract
    # is unchanged — the E2E mock keys on the marker's presence).
    assert build_deflect_prompt(["T1", "T2"]) == (
        _base("LOW")
        + "\nDEFLECT_MODE: retrieval was weak — the titles below are the closest "
        "your notes come to the question. They are titles only; do not pretend "
        "they answer it. Use them to propose 2-3 alternative questions.\n"
        "Reply in plain text only — you have no tools in this mode.\n"
        + "- T1\n- T2"
    )
    assert "<tuning>" not in build_high_prompt([doc])
@@ -299,11 +308,14 @@ def test_no_overview_prompt_is_byte_identical_to_pre_phase() -> None:
    docs_block = "\n<documents>\n" + block + "\n</documents>" + "\n" + TOOLS_SECTION
    high_plain = _base("HIGH") + docs_block
    high_steered = _base("HIGH") + "\n" + build_steering_section(["be concise"]) + docs_block
    # Phase 71: the owner-permitted plain-text line is part of the
    # DEFLECT_MODE body in every LOW build (with or without steering).
    low_plain = (
        _base("LOW")
        + "\nDEFLECT_MODE: retrieval was weak — the titles below are the closest "
        "your notes come to the question. They are titles only; do not pretend "
        "they answer it. Use them to propose 2-3 alternative questions.\n"
        "Reply in plain text only — you have no tools in this mode.\n"
        + "- T1\n- T2"
    )
    low_steered = (
@@ -313,6 +325,7 @@ def test_no_overview_prompt_is_byte_identical_to_pre_phase() -> None:
        + "\nDEFLECT_MODE: retrieval was weak — the titles below are the closest "
        "your notes come to the question. They are titles only; do not pretend "
        "they answer it. Use them to propose 2-3 alternative questions.\n"
        "Reply in plain text only — you have no tools in this mode.\n"
        + "- T1\n- T2"
    )
    for kb in (None, "", "   \n\t "):
@@ -399,3 +412,67 @@ def test_prompt_kb_section_over_settings_budget_capped_with_marker(
        close = section.index("</knowledge_base>")
        section = section[: close + len("</knowledge_base>")]
        assert len(section) <= 200
 # ---------- phase 71: the deflection plain-text line (prevention) ----------
 #: The owner-permitted (2026-09-03) line appended to the ``DEFLECT_MODE``
 #: body — the LOW prompt's only phase-71 change. The E2E mock keys on
 #: the ``DEFLECT_MODE`` marker's *presence*, not the wording, so the
 #: marker-keying contract is unchanged by the appended line.
 PLAIN_TEXT_ONLY_LINE = "Reply in plain text only — you have no tools in this mode."
 def _pre_phase71_low_body() -> str:
    """The ``DEFLECT_MODE`` body exactly as it was before phase 71."""
    return (
        "DEFLECT_MODE: retrieval was weak — the titles below are the closest "
        "your notes come to the question. They are titles only; do not pretend "
        "they answer it. Use them to propose 2-3 alternative questions.\n"
    )
 def test_low_prompt_is_pre_phase_plus_exactly_the_plain_text_line() -> None:
    """Diff pin: the LOW prompt = pre-phase text + exactly the one new
    line, appended to the ``DEFLECT_MODE`` body; the weak-hit title list
    follows exactly as before (and the line occurs exactly once)."""
    prompt = build_deflect_prompt(["T1", "T2"])
    assert prompt == (
        _base("LOW")
        + "\n"
        + _pre_phase71_low_body()
        + PLAIN_TEXT_ONLY_LINE
        + "\n"
        + "- T1\n- T2"
    )
    assert prompt.count(PLAIN_TEXT_ONLY_LINE) == 1
    assert prompt.endswith("- T1\n- T2")  # the title list is untouched
 def test_low_prompt_carries_the_line_and_keeps_the_mock_marker() -> None:
    """The new line is present in the LOW prompt (inside the
    ``DEFLECT_MODE`` body, after the marker) and the ``DEFLECT_MODE``
    marker the E2E mock keys on stays put."""
    for titles, tail in ((["T1"], "- T1"), ([], "(nothing close at all)")):
        prompt = build_deflect_prompt(titles)
        assert "DEFLECT_MODE" in prompt
        assert PLAIN_TEXT_ONLY_LINE in prompt
        assert prompt.index("DEFLECT_MODE") < prompt.index(PLAIN_TEXT_ONLY_LINE)
        # The title list (or the no-titles fallback) follows the line
        # exactly as before.
        assert prompt.endswith(tail)
 def test_plain_text_line_never_leaks_into_high_prompt() -> None:
    """The line is the LOW prompt's: every HIGH build (with/without
    steering/overview) is unchanged and carries none of it."""
    doc = _doc("kubernetes.md", "Talos Linux on three nodes.", "Kubernetes Homelab Cluster")
    for notes, kb in (
        (None, None),
        (["be concise"], None),
        (None, OVERVIEW),
        (["be concise"], OVERVIEW),
    ):
        high = build_high_prompt([doc], notes=notes, kb_overview=kb)
        assert PLAIN_TEXT_ONLY_LINE not in high
        assert "you have no tools" not in high
@@ -0,0 +1,280 @@
 """Unit tests: the deterministic tool-scaffolding filter (phase 71, task 01).
 The matrix from the phase overview: a span in one chunk (exact
 ``stripped_chars``), the observed incident span split across chunks at
 **every** boundary offset of the start token (0..len) plus a mid-span
 and an end-token split, multiple spans in one chunk, the standalone
 sibling tokens (alone, embedded, char-by-char), look-alikes that must
 NOT be stripped, partial markers at EOF (``flush`` emits as-is), and
 exact preservation of surrounding text.
 """
 from __future__ import annotations
 import pytest
 from app.rag.scaffolding import SCAFFOLD_PATTERNS, ScaffoldingFilter
 # The observed incident text (2026-09-03 — the deflected round that
 # streamed raw model tool-scaffolding into the UI).
 SPAN_START = "<|tool_call_start|>"
 SPAN_END = "<|tool_call_end|>"
 SPAN_CONTENT = "[read(path='/homelab/backup-notes.md')]"
 SPAN = f"{SPAN_START}{SPAN_CONTENT}{SPAN_END}"
 TOOL_CALLS = "<|tool_calls|>"
 TOOL_CALL = "<|tool_call|>"
 def _run(text: str, chunks: list[str] | None = None) -> tuple[str, ScaffoldingFilter]:
    """Feed *chunks* (or *text* as one chunk) + ``flush``; return the
    (total output, filter) pair the assertions below work on."""
    f = ScaffoldingFilter()
    parts = [f.feed(chunk) for chunk in (chunks if chunks is not None else [text])]
    parts.append(f.flush())
    return "".join(parts), f
 # ---------------------------------------------------------------- the registry
 def test_registry_is_the_three_observed_forms() -> None:
    # Every entry traces to the 2026-09-03 incident — no speculative
    # entries (the registry is the extension point; new forms need an
    # observed capture + a fixture here).
    assert len(SCAFFOLD_PATTERNS) == 3
    assert SCAFFOLD_PATTERNS[0].pattern == r"<\|tool_call_start\|>[\s\S]*?<\|tool_call_end\|>"
    assert SCAFFOLD_PATTERNS[1].pattern == r"<\|tool_calls\|>"
    assert SCAFFOLD_PATTERNS[2].pattern == r"<\|tool_call\|>"
    # The span form is non-greedy: two spans each strip to their own end.
    assert SCAFFOLD_PATTERNS[0].sub("", SPAN + "x" + SPAN) == "x"
 # ---------------------------------------------- span in one chunk, split spans
 def test_span_in_one_chunk_is_stripped_with_exact_count() -> None:
    out, f = _run(SPAN)
    assert out == ""
    assert f.stripped_chars == len(SPAN)
 def _split_offsets() -> list[int]:
    """Every split offset of the start token (0..len) + a mid-span split
    and an end-token split (the task-01 matrix)."""
    offsets = list(range(len(SPAN_START) + 1))  # 0 .. 19, every start-token offset
    offsets += [
        len(SPAN_START) + 1,  # one char into the span body
        len(SPAN_START) + len(SPAN_CONTENT) // 2,  # mid-span
        len(SPAN) - len(SPAN_END),  # exactly at the end token
        len(SPAN) - 5,  # inside the end token
        len(SPAN) - 1,  # the very last char
    ]
    return offsets
@pytest.mark.parametrize("offset", _split_offsets())
 def test_span_split_at_boundary_offsets_emits_nothing_until_complete(offset: int) -> None:
    f = ScaffoldingFilter()
    first = f.feed(SPAN[:offset])
    second = f.feed(SPAN[offset:])
    third = f.flush()
    assert first == ""  # nothing ever emits until the span completes
    assert second == ""
    assert third == ""
    assert f.stripped_chars == len(SPAN)
 def test_span_completed_by_a_later_chunk_then_content_follows() -> None:
    f = ScaffoldingFilter()
    assert f.feed(f"lead {SPAN_START}part") == "lead "  # prose before the span flows
    assert f.feed(f"{SPAN_END} tail") == " tail"
    assert f.flush() == ""
    assert f.stripped_chars == len(SPAN_START) + 4 + len(SPAN_END)  # body: "part"
 def test_span_with_nested_start_token_strips_to_first_end() -> None:
    # Non-greedy: the span runs from the FIRST start to the FIRST end.
    text = f"{SPAN_START}a{SPAN_START}b{SPAN_END}"
    out, f = _run(text)
    assert out == ""
    assert f.stripped_chars == len(text)
 def test_spans_across_chunks_with_a_standalone_token_after() -> None:
    f = ScaffoldingFilter()
    assert f.feed(SPAN[:25]) == ""  # inside the span body
    assert f.feed(SPAN[25:] + " mid " + TOOL_CALLS) == " mid "
    assert f.flush() == ""
    assert f.stripped_chars == len(SPAN) + len(TOOL_CALLS)
 # ------------------------------------------------------- multiple spans / tokens
 def test_two_spans_in_one_chunk_strip_each_to_their_own_end() -> None:
    out, f = _run(f"{SPAN} middle {SPAN}")
    assert out == " middle "
    assert f.stripped_chars == 2 * len(SPAN)
 def test_two_spans_with_different_content() -> None:
    other = f"{SPAN_START}[grep(pattern='vault')]{SPAN_END}"
    out, f = _run(f"before{SPAN}mid{other}after")
    assert out == "beforemidafter"
    assert f.stripped_chars == len(SPAN) + len(other)
 def test_repeated_strip_until_no_complete_match_remains() -> None:
    # A standalone token embedded in a would-be token: step (a) repeats
    # leftmost-complete removals until the buffer is clean.
    out, f = _run(f"x<|tool_call{TOOL_CALLS}|>y")
    assert out == "xy"
    assert f.stripped_chars == len(TOOL_CALLS) + len(TOOL_CALL)
 # ------------------------------------------------------------ standalone tokens
@pytest.mark.parametrize(
    ("token", "expected"),
    [(TOOL_CALLS, 14), (TOOL_CALL, 13)],
 )
 def test_standalone_token_alone_is_stripped(token: str, expected: int) -> None:
    out, f = _run(token)
    assert out == ""
    assert f.stripped_chars == expected
@pytest.mark.parametrize("token", [TOOL_CALLS, TOOL_CALL])
 def test_standalone_token_embedded_in_a_line_is_stripped(token: str) -> None:
    out, f = _run(f"line {token} tail\n")
    assert out == "line  tail\n"
    assert f.stripped_chars == len(token)
 def test_standalone_tokens_adjacent_to_text() -> None:
    out, f = _run(f"a{TOOL_CALL}b{TOOL_CALLS}c")
    assert out == "abc"
    assert f.stripped_chars == len(TOOL_CALL) + len(TOOL_CALLS)
@pytest.mark.parametrize("token", [TOOL_CALLS, TOOL_CALL])
 def test_standalone_token_char_by_char_never_emits(token: str) -> None:
    out, f = _run("", list(token))
    assert out == ""
    assert f.stripped_chars == len(token)
 def test_partial_standalone_token_completing_in_a_later_chunk() -> None:
    f = ScaffoldingFilter()
    assert f.feed("<|tool_ca") == ""
    assert f.feed("lls|>") == ""
    assert f.flush() == ""
    assert f.stripped_chars == len(TOOL_CALLS)
 def test_partial_token_resolving_to_prose_is_emitted() -> None:
    # ``<|tool_cat|>`` is no known form — the held partial prefix must be
    # released as content once the input can no longer complete a token.
    f = ScaffoldingFilter()
    assert f.feed("x" * 50 + "<|tool_ca") == "x" * 50  # live prefix held
    assert f.feed("t|> done") == "<|tool_cat|> done"
    assert f.flush() == ""
    assert f.stripped_chars == 0
 # ------------------------------------------------------------------- look-alikes
@pytest.mark.parametrize(
    "text",
    [
        "tool_call_start",  # the word as prose, no delimiters
        "the tool_call and tool_calls words",  # prose words
        "<|tool_call_start|",  # missing delimiter close — prose at EOF
        "<|some_other_token|>",  # unknown token
        "<|tool_call_end|>",  # a lone end token without a start
        "a < b",  # a lone angle bracket
    ],
 )
 def test_lookalikes_are_not_stripped_and_emit_verbatim(text: str) -> None:
    out, f = _run(text)
    assert out == text
    assert f.stripped_chars == 0
 def test_lone_end_token_before_an_open_start_stays_content_until_closed() -> None:
    # The lone end token is content (emitted as-is); the trailing start
    # opens a span and is HELD — if the span later closes, only the
    # span is stripped, the lone end token stays.
    f = ScaffoldingFilter()
    assert f.feed(f"{SPAN_END}{SPAN_START}") == SPAN_END
    assert f.feed(f"body{SPAN_END}") == ""
    assert f.flush() == ""
    assert f.stripped_chars == len(SPAN_START) + 4 + len(SPAN_END)
 def test_lone_end_token_at_eof_flushes_as_is() -> None:
    f = ScaffoldingFilter()
    out = f.feed(SPAN_END)
    assert out + f.flush() == SPAN_END  # verbatim total (the tail may be split)
    assert f.stripped_chars == 0
 # ----------------------------------------------------------- EOF / empty / clean
 def test_partial_start_token_at_eof_flushes_as_is() -> None:
    # A partial marker at EOF is content, not scaffolding (pinned choice).
    f = ScaffoldingFilter()
    assert f.feed("<|tool_call_st") == ""
    assert f.flush() == "<|tool_call_st"
    assert f.stripped_chars == 0
@pytest.mark.parametrize(
    "partial",
    [
        "<",
        "<|",
        "<|tool_ca",
        "<|tool_call|",  # one char short of the standalone token
        "<|tool_calls|",  # one char short of the standalone token
        "<|tool_call_start|",  # one char short of the span opening
    ],
 )
 def test_partial_tokens_at_eof_flush_as_is(partial: str) -> None:
    out, f = _run(partial)
    assert out == partial
    assert f.stripped_chars == 0
 def test_empty_chunk_is_a_noop() -> None:
    f = ScaffoldingFilter()
    assert f.feed("") == ""
    assert f.feed("") == ""
    assert f.flush() == ""
    assert f.stripped_chars == 0
 def test_empty_chunk_amid_pending_is_a_noop() -> None:
    f = ScaffoldingFilter()
    assert f.feed("<|tool_ca") == ""
    assert f.feed("") == ""
    assert f.feed("lls|>") == ""  # completes the standalone token
    assert f.flush() == ""
    assert f.stripped_chars == len(TOOL_CALLS)
 def test_flush_on_a_clean_stream_emits_nothing_new() -> None:
    f = ScaffoldingFilter()
    assert f.feed("hello ") == "hello "
    assert f.feed("world") == "world"
    assert f.flush() == ""
    assert f.stripped_chars == 0
 def test_surrounding_text_preserved_exactly() -> None:
    # No reflow beyond the removal — both spaces around the span stay.
    out, f = _run(f"hello {SPAN} world")
    assert out == "hello  world"
    assert f.stripped_chars == len(SPAN)