Phase 75 (TODO.md L4): "Save as doc" now drafts a document from the
ENTIRE chat session — every question and answer up to the click, in
order — instead of only the clicked bubble's answer; the existing
doc-edit screen's free-form body editing is how the user edits out
anything they don't want to keep from previous replies (no new UI
surface).
Task 01 (frontend):
- app.js buildSessionTranscript(): walks the bor.chat.v1 conversation
record in order — a numbered section per user turn ("## N.
<question, raw>" + blank line + the raw answer text; more answers
join under the same heading), sections blank-line separated, all
trailing whitespace collapsed to one final newline. Only the raw
persisted text travels (m.who + m.text — no thinking blocks, no
source chips, no tune metadata); a brain record before the first
user record is skipped; a heading-only section marks a user turn
whose answer never landed (A6, owner-confirmed 2026-09-08).
- saveAsDoc(btn): the draft body is buildSessionTranscript(); the
dead single-bubble markdown parameter is dropped (the button's
appendSaveAsDocButton signature is unchanged — one button per
bubble). Title/path/double-click guard/hand-off are unchanged
(defaultDocTitle: the last question, whitespace-collapsed,
<=120 chars; docs/<slug>.md).
- Unit: the app.js source pins move to the transcript shape (whole
session, no thinking, no dead parameter).
Task 02 (E2E):
- tests/e2e/test_save_doc_session.py (bare-repo fixture, the
phase-59 convention — git as source of truth): three DISTINCT
on-topic turns in one session (turn 1 carries the phase-17
"think out loud" trigger so its record has a thinking block the
transcript must exclude) -> save on the LAST bubble -> the
prefilled body is ## 1./## 2./## 3. in order, byte-exact against
the deterministic mock, thinking-free -> edit the whole
section-2 block out of the body -> push -> git show
bor-docs:<path> equals the EDITED body byte-for-byte (section 2's
question and answer provably absent; sections 1 and 3 byte-exact;
the UI's sha prefix is git rev-parse bor-docs). Second test:
the button on the FIRST bubble still drafts the whole session
(A6 — the transcript is the session at click time, title stays
the last question); canceling leaves the branch tip untouched.
- tests/e2e/test_response_to_docs.py: the phase-59 single-turn body
expectation moves to the transcript shape ("## 1. <question>" +
the answer's markdown) — the rest of the suite unchanged.
Also lands the phase-74 file moves (00_phase.md /
03_mock_marker_e2e.md -> complete/) and the phase reports — the
house convention of committing .agents/ with the phase.
Phase 74 (TODO.md L4): a follow-up question now reaches the model WITH
the conversation so far — every prior user/brain turn and the prior
thinking blocks on brain turns (preserve-thinking) — while
POST /api/chat stays stateless (A10): the client provides the history
in the request body and the server stores nothing new.
Server (task 01):
- ChatRequest.history: optional list[HistoryTurn] (who: user|brain,
text, optional thinking) — absent/empty keeps the request
byte-identical to pre-phase-74 (the two-message [system, user]
request; the kill-switch semantics are pinned in the integration
suite).
- app.rag.prompts.history_to_messages: pure mapper — walks the turns
newest-first against the settings budgets (history_max_turns=40 /
history_max_chars=24000, BOR_HISTORY_MAX_TURNS /
BOR_HISTORY_MAX_CHARS); a capped turn is dropped WHOLE (never cut
mid-answer); the kept window is returned oldest-first; brain turns
carry their thinking as reasoning_content (A4) only when
non-empty.
- Both branches feed it: the deflected path splices it between the
system prompt and the current user message (the phase-71 recovery
still rebuilds from messages[1:]), the grounded agent receives
run_agent(..., history=hist); llm.py's message params widen to
list[dict[str, Any]] (string-only messages stay byte-identical on
the wire — the SDK passes message dicts through verbatim).
- The per-turn log line (PLAN §9) gains history_msgs=N after
kb_chars=N.
- Pins: tests/unit/test_history.py (mapper: mapping, reasoning
gating, both budgets, drop-whole, ordering, empty default),
tests/unit/test_config.py (the two settings + env overrides),
tests/unit/test_agent.py (the history splice + the default),
tests/integration/test_chat_api.py (deflected AND grounded forward
the history incl. reasoning_content, no-history byte-identity, 422
pins, the log field).
Client (task 02):
- runTurn — the single funnel for fresh send / phase-49 retry /
phase-53 stale-regen — sends history = the conversation record
minus the current question, with thinking only on brain records
that streamed one (undefined drops the key from the JSON, the
record's convention); the question is never duplicated into the
history.
Wire proof (task 03):
- The mock's echo my history marker (HISTORY_TRIGGER) answers with
the deterministic history echo — history: N prior messages; last
answer tail: <last 24 chars>; thinking: yes|no — checked BEFORE
the DEFLECT_MODE branch (like TABLE_TRIGGER), so it fires on both
turn branches whatever the gate says; the module docstring records
the user/assistant-only history invariant that keeps every
existing (tool-result-classified) marker flow unaffected.
- tests/e2e/test_llm_history.py (isolated): a grounded follow-up and
a deflected follow-up both receive history: 2 prior messages +
thinking: yes + the byte-exact tail of turn 1's answer (derived
from the persisted bor.chat.v1 record — the same array the client
maps into the body); a cold start receives history: 0 prior
messages / last answer tail: none / thinking: no.
- Regressions green in isolation: chat_rag, chat_history (phase 50),
agent_document_tools, harness_aligned_tools, stop_generation,
retry_answer, response_to_docs.
Root cause (task 01): none of C1-C3 - in Chromium 151 (real mode) a
merely-hidden tab neither stops the stream (frames arrive at full rate;
turn completes) nor fires pagehide on tab switch; C1's double-record
path was proven latent via a synthetic pagehide (trigger is
browser-dependent, e.g. Safari) and C2 (the 120s pre-token guard) was
confirmed to fire while hidden.
- C1: the pagehide partial-persist is correlated with the turn's settle
(leavePartialIndex) - the done/stop settle REPLACES it in place
(identity-guarded rememberBrainTurn in-place mode), so bor.chat.v1
and the auto-saved saved_chats row keep exactly ONE brain turn per
question; a real navigation never runs a settle, so the leave-save
is unchanged.
- C2: the visibility re-arm gives the still-armed pre-token guard a
fresh TURN_TIMEOUT_MS when the tab returns to visible - hidden time
no longer counts toward the 120s guard.
- Phase-48 teardown contract untouched: Stop / tab close / real
navigation still cancel the fetch and stop the model.
- Unit pins: tests/unit/test_frontend_hidden_tab.py (the app.js
mechanisms without a browser).
- E2E pins: tests/e2e/test_hidden_tab_stream.py - synthetic pagehide
mid-stream completes exactly once with one brain turn (localStorage
+ auto-saved row), reload restores one bubble, no-event baseline,
and the fake-clock pre-token guard re-arm (discriminating: fails
with the re-arm disabled).
Convert the two unchecked TODO items into executable phases (Protocol B,
appended after the 72 completed phases):
- 73_hidden_tab_stream (TODO L3): a merely-hidden browser tab must never
stop a generating answer; repro/root-cause decision tree + the pagehide
partial-correlation fix + the hidden-tab E2E pin.
- 74_llm_chat_history (TODO L4, history): client-provided history in
POST /api/chat (stateless, A10) mapped through both the deflected and
grounded agent paths, prior thinking blocks preserved via
reasoning_content, capped oldest-first; mock echo marker + E2E.
- 75_save_doc_full_session (TODO L4, save-as-doc): the Save-as-doc draft
body becomes the full session transcript; edit-out happens in the
existing doc-edit body; multi-turn git-verified E2E.
Owner-confirmed assumptions A1-A7 are recorded as ASSUMPTION lines in the
task files. TODO.md is cleared (items now live in .agents/phases/todo/).
Standardize on the .agents/ directory (shared with project skills):
phases/, user_stories/, reports/, screenshots/, validate.sh, and
phase-sessions/ + pipeline.log all move to .agents/ (git mv preserves
history; runtime artifacts move alongside).
Updates every reference in AGENTS.md, README.md, .gitignore, app
docstrings, and test story headers. Historical KB content in data/
and the runtime pipeline.log transcript are left untouched.
Codifies the 2026-09-05 turbo comparison workflow as a project skill under
.agents/skills/: switch BOR_LLM_CHAT_MODEL in .env, run the fixture gate
(twice, for variance) + the locked derived gate with per-turn wall timing,
interpret the two metrics against the reference model rates (re-read habit:
lite ~100%, turbo ~12%; usage-floor MISS as test artifact; caps as real
regression), record the verdicts byte-exact in TOOL_CALLING_TESTING.md, and
commit the doc. Rules baked in: never touch the battery/thresholds/fixtures,
never edit app code, never commit .env.
turbo (2026-09-05, same fixture KB): fixture gate PASS 100%/100% on both
metrics, two runs (wall 105-135s vs lite 43-55s); the redundant re-read
of seeded documents that capped lite's executed ratio at 58-73% is
model-specific (turbo re-read rate ~12% vs ~100% in-sample), corroborating
section 7's framing. Locked derived battery: turbo fails only the >=6/10
tool-turn usage floor (it answers seeded read-target questions from
context instead of making the refusable read call) - accuracy on all
emitted calls still 100%/100%.
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 phase-72 iteration loop cleared the database, git-cloned the homelab repo, re-imported 38-51 documents and re-embedded per run — many minutes per iteration against a different KB every time (owner directive 2026-09-04: stop importing the homelab repo on every test run). Replace it with:
- tests/fixtures/agent_kb/: 8 hand-written markdown docs (sources 'deployments'/'homelab') whose specifics (rack7, 10.77.42.0/24, VLAN 130, rbm-8842, 17 2 * * *, obsidian-bor:2026.7.14, 18765, 18443, ...) no model can guess; read targets carry non-topical filenames so their questions do not lexically seed them (the read must actually happen)
- tests/fixtures/test_kb.dump.sql: data-only snapshot (TRUNCATE + INSERTs incl. embeddings, self-contained git_sources rows, static KB overview) — verified by round-trip checksum at build time
- scripts/load_test_kb.py: one-off rebuild (real pipeline + embeddings, ~2s) that also prints the per-question retrieval report (all 10 battery questions must be grounded)
- scripts/restore_test_kb.py: sub-second one-transaction restore (no git clone, no re-embedding)
- scripts/agent_realmodel_check.py: the gate gains --restore / --mode fixture (curated 10-question battery with one unambiguously correct tool behavior per question) / --turns N (12s micro-loop) / --concurrency / per-turn + total wall timing, and a second accuracy metric (contract accuracy: well-formed calls targeting resolvable entities) alongside the phase-72 locked executed ratio — the re-read of a seeded doc is a copy-invariant model behavior (5 variants, 0/15 flipped) that the dedupe refusal counts as a failure
- TOOL_CALLING_TESTING.md: the human-readable methodology (fast loop, design rules, metrics, copy levers + tried-and-reverted table, current standing, open design question)
Measured: restore 0.03s; micro-loop ~12s; full loop ~43-55s; concurrency 2/3 gives no gain (endpoint serializes).
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.
meta description -> locked (A3) auto-save string (history.html L6).
page-sub -> locked (A3) string, the <strong>Save</strong> emphasis retired with the button (L106-109).
empty row -> locked (A3) string; colspan=6, hidden, and row id untouched (L163).
h1, the anonymous gate section, and history.js are byte-identical — state language verified accurate.
- task 01: relocate the .chat-actions row (New chat + Share, comments byte-identical with a Phase 65 note) from the top of the column to the bottom of .chat-shell, directly above the composer
- task 02 (owner-locked A1): wrap the row + #composer in ONE sticky .chat-bottom unit (position: sticky; bottom: env(safe-area-inset-bottom, 0), no z-index) — the pills stay at the bottom of the screen at every scroll position and settle into flow above the footer
- task 03 (owner-locked A2): right-align the bottom row to the column's right edge (justify-content: flex-end), mirroring the right-aligned Save-as-doc corner; the five action pills share one 44px / 999px-pill geometry
- task 04: dedicated Playwright suite tests/e2e/test_bottom_chat_actions.py (resting geometry, the A1 pin across the sticky range, A2 alignment + DOM order + mobile stack + 360px overflow bound + 44px touch targets, New chat / Share click-through) — green in isolation
- task 05: regression matrix green in isolation (pinned_composer 4, save_share_ux 5, chat_persistence 4, share_chat 4, chat_history 5, smoke 3); full gate green — unit + integration pass, app/ coverage 99% (>90%), ruff + pyright clean
Phase 65 (TODO.md L3): move the New chat + Share cluster to the pinned
bottom of the chat column (sticky .chat-bottom unit with the composer)
and right-align the row to the Save-as-doc action corner.
Phase 66 (TODO.md L4): History tab copy — every chat saves
automatically; retire the Save-button references.
Fixed: index.html meta description, empty-state sub and composer
placeholder (A1); app/config.py default suggestion chips → the four
neutral A2 defaults (BOR_SUGGESTIONS override unchanged); sources.html
KB page-sub → the current source model (git repos + local dirs +
uploaded archives, Sync pulls/imports); git-sources.html example URL
→ your-repo.git (A3); all 9 footers → neutral default in
span.footer-text (the phase-62 hook); E2E/unit conftests force the
code defaults so a local .env cannot leak corpus copy into tests;
new unit text pins + dedicated E2E suite.
Task 02 verification read-through — no change needed:
- sources.html sync result/error copy (matches the real sync behavior)
- tuning.html page-sub (accurate as written)
- history.html page-sub (accurate as written)
- doc-edit.html page-sub (accurate as written)
- git-sources.html page-sub (accurate as written)
- #sources-gate anonymous copy (accurate as written)
Remove the blanket .agent/ gitignore so the phase roadmap, user
stories, reports, and PLAN.md are versioned with the code. Only
runtime artifacts (.agent/phase-sessions/, .agent/pipeline.log)
remain ignored. Update AGENTS.md git protocol rule to match.
Phase 52's first pass shipped `position: sticky; bottom` on `.composer` and
called the phase done, but the owner's requirement — "the chat message-input
textarea should be at the bottom of the screen" — still failed in the browser:
on an empty/short chat the input rested just under the empty state (~57% of
the viewport) with a dead band down to the footer.
`position: sticky` can only pull a box UP toward the scrollport's bottom edge;
it can never push a box DOWN to meet it, so on a page that does not overflow
it is a no-op. The old story suite only exercised an overflowing conversation
(one test even asserted the buggy resting position as expected), which is why
the half-fix passed.
- `.messages { flex: 1 1 auto }` — absorbs a short page's free space so the
composer's resting in-flow position is the bottom of the full-height column
(body min-height:100dvh -> .app-main flex:1 -> .chat-shell flex:1); basis
stays `auto`, no height cap, no overflow — the document stays the scroller
- `.composer { bottom: env(safe-area-inset-bottom, 0) }` — the explicit 0
fallback replaces the env()-only offset, which degraded to `auto` (no pin)
wherever env() is unsupported
- E2E: `test_empty_chat_composer_sits_in_normal_flow` ->
`..._at_the_screen_bottom` (chrome-only band below the resting composer);
the phone suite now checks the resting position as well as the pinned one
- Unit pins: the flex-grow half and the full-height column are pinned, so the
fix cannot silently regress to sticky-only
Still CSS-only — no DOM change, no JS, no new scroll call site (phase 42
never-auto-scroll contract intact), no z-index.
Verified: 1019 unit/integration tests pass (app/ coverage 99%), ruff and
pyright clean; tests/e2e/test_pinned_composer.py green in isolation (4), plus
the stop/autoscroll/persistence/mobile-nav suites and 14 layout/scroll
neighbours green in isolation.
- scrollReveal lands at the document bottom (window.scrollTo) instead of
scrollIntoView({ block: 'end' }): the old alignment sat above the
in-flow composer, so every Enter hopped the page up by the
composer+footer height and pushed the composer below the fold.
- The thinking window's pin state is now captured BEFORE the re-render
(const pinned = block.open && isThinkingNearBottom(textEl)): the
post-render distance read the new chunk's rendered height, not the
user's position, so any chunk taller than the 32px band (real-model
deltas, '\n\n' paragraph breaks) killed the follow at the first
2-newline gap.
- Mock LLM: new 'think in paragraphs' trigger (scratchpad with real
blank-line breaks, 60-char frames) — the 12-char mock frames never
rendered past the band, which is why the bug survived the E2E gates.
- E2E (both verified red against the old code):
test_submit_does_not_hop_up, test_thinking_window_follows_across_paragraph_breaks.
- Unit source-marker tests updated to the new contracts.
Phase 49 (owner request, chat 2026-08-28: "The git sources page should
remove local directory and should instead accept a tarball or zipfile
upload which it will unpack and scan … reuploading the same tarball
should not create a new folder, but should unpack and overwrite the
previously unpacked content" — design confirmed in the same
conversation):
* POST /api/git-sources/upload (admin-only, require_admin): accepts
.tar/.tar.gz/.tgz/.zip, streams it with the BOR_UPLOAD_MAX_MB cap
(bounds BOTH the compressed upload and the total extracted bytes —
zip-bomb guard), safely unpacks (absolute/traversal/symlink/hardlink
escape and device/FIFO members rejected), and atomically swaps the
content in over BOR_UPLOAD_DIR/<name>/ (name = filename minus the
archive suffix — no missing window, a failed upload never touches the
existing folder/row/KB). The git_sources row is upserted by path
(kind='local', no duplicates, added_at preserved), the models are
checked fail-fast (503 sanitized when down — the folder/row stay
committed and the next sync/re-upload retries idempotently), and the
source is scanned synchronously in the request (single-source
import_sources prune=True + change-gated KB overview), answering 200
with the sync-style counts. One upload at a time (409); the request
session is released before the scan so a concurrent TRUNCATE cannot
deadlock against it.
* app/rag/archive_upload.py: ArchiveUploadError, ARCHIVE_SUFFIXES,
archive_source_name (safe-name derivation), unpack_archive (guarded
zip/tar extraction with the extracted-byte cap, no partial state),
swap_in (atomic replace with restore-on-failure) — fully unit-tested.
* app/config.py + .env.example: BOR_UPLOAD_DIR (default
~/bor-sources/uploads, deliberately separate from the git checkouts)
and BOR_UPLOAD_MAX_MB (default 512; a validator fails loud at
startup on <= 0).
* python-multipart added to the dependencies — FastAPI's required
multipart parser (an A2 implementation detail, phase locked decision).
* The Sources page: the phase-38 "Add a local directory" form is
removed; #archive-upload-form takes its place (labeled file input,
"Upload & scan" button, the §7.4 never-stale lifecycle, inline
role=alert error, role=status count line); hint + table caption
updated. The POST /api/git-sources kind=local API contract is
UNCHANGED — a plain directory is still registrable via the API, and
existing Local rows list/remove/sync exactly as before.
* The phase-38 story E2E (test_local_directory_sources.py) is rewritten
API-driven — the form it drove is gone; its acceptance stands.
* The story E2E (test_archive_upload_sources.py): the swap,
upload→scan→list (the deterministic "Uploading…" in-flight state, the
Local row, /api/docs + the RAG catalog), same-filename re-upload
(in-place replace, prune, no duplicate row, v2-only folder), the
422 inline error + recovery (the form is not wedged), and the
anonymous gate + 403.
* README: the archive-upload section (formats, naming rule, in-place
replace, both new settings), the local-directory form removal noted,
config reference rows for BOR_UPLOAD_DIR / BOR_UPLOAD_MAX_MB.
Gates: unit+integration green, app/ coverage 99%, the story E2E green
in isolation, the regression suites (git sources admin, local
directory sources, sync button, import documents, nav rename, smoke,
shared header) green in isolation, ruff + pyright clean.
Note: per this phase's file-level staging, frontend/assets/styles.css
also carries the small same-day in-flight owner rework already in the
working tree (the .sign-in-mobile companion rule for the phase-48
mobile sign-in copy); the phase-49 change is the upload form's block.
ducoterra