Phase 72 (72_teaching_refusals) — completed under the 2026-09-04 controlled
methodology (owner directive: stop clearing/re-importing the homelab KB per
iteration; measure tool-calling accuracy on a controlled fixture KB, target
>90%).
Real-model gate verdicts (live, configured chat model 'lite', fixture KB):
- Controlled fixture battery (the new methodology's pass condition —
contract accuracy >= 90%): PASS, 4 consecutive runs:
gate: lite PASS turns=10 answered=10 caps=0 tool-turns=10 calls 8/11 executed (73%) contract 11/11 (100%) 2026-09-04 (wall 43.4s)
gate: lite PASS turns=10 answered=10 caps=0 tool-turns=10 calls 8/13 executed (62%) contract 12/13 (92%) 2026-09-04 (wall 50.6s)
gate: lite PASS turns=10 answered=10 caps=0 tool-turns=10 calls 7/11 executed (64%) contract 11/11 (100%) 2026-09-04 (wall 46.8s)
gate: lite PASS turns=10 answered=10 caps=0 tool-turns=10 calls 9/15 executed (60%) contract 14/15 (93%) 2026-09-04 (wall 54.8s)
- Locked derived battery (phase-72 task 05, executed >= 90% bar, run
unchanged on the same fixture KB):
gate: lite FAIL turns=10 answered=10 caps=0 tool-turns=10 calls 5/15 executed (33%) contract 12/15 (80%) 2026-09-04 (wall 47.7s)
The teaching works — every bare-path trap self-corrects in exactly one
round, zero cap hits, zero repeat loops, 10/10 answered. The locked
executed bar is blocked by ALREADY_IN_CONTEXT dedupe refusals on the
corrected re-reads (the trap question seeds its target, so the correct
combined-form read is refused for redundancy) — a copy-invariant model
behavior (five copy variants, 0/15 re-reads flipped, 2026-09-03 -> 04)
and an app-semantics decision for the owner (TOOL_CALLING_TESTING.md
sections 5 and 7), not a copy lever.
Copy changes this phase owns (unit pins updated to follow):
- app/rag/agent.py: ls teaching refusals (path-like scope -> document-path
line; unknown source -> no-source line with the source-name
parenthetical), read/grep 'did you mean source/path?' teaching
(find_path_candidates: exact or suffix path match, catalog order, cap 3),
ALREADY_IN_CONTEXT naming the correct action (answer from the text
already in the prompt), read tool description front-loaded with the
do-not-read rule (the 2026-09-04 controlled telemetry: the re-read is
the only remaining refusal class; contract accuracy 92-100% across runs)
- app/rag/prompts.py: TOOLS_SECTION states the document-identity contract
up front (ls path = source name; read/grep = combined source/path
including the source name; do-not-read for <documents> documents placed
next to the read teaching; one-call-per-reply and never-repeat rules)
- tests: refusal pins (unit + integration), new dedicated E2E suite
tests/e2e/test_tool_path_teaching.py (mock misuse flow, green in
isolation), regression suites green in isolation (harness_aligned_tools,
agent_document_tools, agent_unlimited_tools, search_tool, chat_rag).
Gates: uv run pytest green (1501); coverage TOTAL 99% (>90%); ruff +
pyright clean. Carries the still-uncommitted phase-71 todo/ -> complete/
move and both phases' .agent/reports/ (AGENTS.md 8).
The model treated the combined 'source/path' string (as printed in
search result lines, read-result headers and refusals) as the
document's identity and passed it as 'source' — e.g.
source='homelab/active/container_caddy/caddy.md' instead of
source='homelab', path='active/container_caddy/caddy.md'.
- Rewrite the read_document description with the split rule (source =
before the FIRST '/', path = after it) and a worked example; share
the source/path parameter descriptions between read_document and
search_documents; map search result lines back onto the split.
- New _resolve_document: on a lookup miss with a '/' in source, retry
at the first slash (source names are directory basenames and can
never contain '/'), plus a continuation candidate for a split at a
later slash; a self-corrected combined form for an already-in-context
document is still rejected as ALREADY_IN_CONTEXT.
- A slash-carrying source that matches nothing gets an educational
refusal naming the corrected arguments instead of the generic line
that repeated the combined form.
Verified live against aipi (lite) + the imported homelab KB: A/B on
the exact failure scenario (5 runs each, right after a
combined-source search result) — old descriptions 5/5 combined, new
descriptions 5/5 clean; two live UI turns (Playwright) produced only
clean split arguments, including a multi-hop read of
install_caddy_deskwork.yaml that landed in done.sources. Full suite:
1376 passed, app coverage 99% (agent.py 100%), ruff + pyright clean,
agent/search E2E green in isolation.
Phase 45 (owner permission 2026-08-27, TODO.md L8: "allow the LLM
to make as many tool calls as it wants"): the phase-37 per-turn tool
budgets (BOR_AGENT_LIST_CALLS / BOR_AGENT_READ_CALLS, default 1 each)
and their exhaustion refusals are removed — a grounded turn now offers
list_documents / read_document for the whole turn (re-lists included),
bounded only by the round cap:
- app/config.py: agent_max_rounds (BOR_AGENT_MAX_ROUNDS, default 10,
negative rejected) replaces agent_list_calls / agent_read_calls;
.env.example + README document the single knob; app/rag/prompts.py
docstrings follow.
- app/rag/agent.py: the loop runs tools until the model answers or
rounds >= max_rounds, at which point it forces one final no-tools
answer (the cap is the only forced exit); 0 = no tools — exactly one
tools=None request, byte-identical to the pre-phase-37 path (the
kill switch). Rejected calls (unknown tool / missing args /
already-in-context / unknown path) still consume a round, so
pathological rejected-call streams are bounded by the cap. The
per-call log line is now tool/args/round=N/M; the per-turn
tool_calls=N field and the tool SSE event are unchanged.
- tests/e2e/mock_llm.py: MULTI_READ_TRIGGER ("read two documents") —
the deterministic list -> read #1 -> read #2 -> forced-answer flow
(byte-stable "I read <sp1> and <sp2>." line), classified by the
count of tool-role read results; the phase-37 single-read flow stays
byte-identical (unit-pinned in tests/unit/test_mock_tool_flow.py).
- tests/e2e/test_agent_unlimited_tools.py (new, story suite,
mock-only): three tool frames/lines in order (one list, two reads —
the second read is what the old read budget refused) + the
both-named non-deflected answer; done.sources + chips = retrieval
doc + both reads, deduped; no budget refusal rendered; the
single-read marker flow regression (exactly one read, single tool
pair).
- .agent/PLAN.md: the phase-45 SSE revision note (owner-locked, R2) —
the only PLAN edit this phase; the phase-37 note's budget clause is
marked removed.
Unit/integration rewrites (test_agent.py round-cap matrix incl. the
kill switch and rejected-call spam, test_config.py, test_chat_api.py
agent_max_rounds=0 fixtures) landed with the server core so every gate
stays green.
uv run pytest: 756 passed, app/ coverage 99%; ruff + pyright clean;
story E2E 4/4 in isolation (ran twice); regression E2E suites
(agent_document_tools unmodified, chat_rag, smoke) green in isolation.
Also records the 45_agent_unlimited_tools todo/ -> complete/ task-file
moves (00/01/02 pending in the working tree, task 03 moves on success).
Grounded chat turns now run the agent loop (app/rag/agent.py) instead
of a bare chat_stream: while the per-turn budgets last
(BOR_AGENT_LIST_CALLS / BOR_AGENT_READ_CALLS, default 1 each) the model
gets list_documents (the indexed catalog, /api/docs order) and
read_document (full text, never truncated — A7-revised contract); once
both budgets are spent the tools key is dropped from the request and
the model must answer. Rejected calls (unknown tool, unknown/missing
path, document already in context, spent budget) consume no budget.
Budgets 0/0 make exactly one tools=None request — byte-identical to
the pre-phase path (budgets-as-kill-switch). Deflected turns keep the
direct chat_stream (A8 unchanged; the LOW prompt never carries the
<tools> section).
SSE contract gains {"type":"tool","name":...,"argument":
"source/path"|null} frames ahead of the answer deltas (PLAN §4
extension, owner permission 2026-08-26); done.sources, query_log.sources
and the per-turn log line (gains tool_calls=N) report the retrieval
docs + read docs, deduped. The UI shows a "calling tool"
button/label state and one visible .tool-call line per call above the
answer; the lines persist with the chat record and re-render on
reload. chat_stream passes tools through and accumulates streaming
tool_calls deltas into ToolCallPiece (tools=None stays byte-identical).
E2E: deterministic mock tool flow ("use your tools" + <tools> marker:
list -> read first catalog line -> quoted answer) plus the story suite
(marker flow, reload re-render, plain/deflected no-tool regressions).
Docs: .env.example + README (the two tools, the budgets, the SSE tool
frame, the "calling tool" UI state).
probe: turbo tool_calls=supported 2026-08-26 (uv run python -m
scripts.llm_probe --tools — non-streaming + streaming
finish_reason=tool_calls, indexed delta.tool_calls partials)
ducoterra