docs+fix: bug_chat_long_conv dogfood artifacts + dashboard truncation fix#73
Merged
Conversation
…n-mode bug_fix.md
Captures the work products from this session's dogfood runs:
* idea.md — /idea-preflight Audit & Patch (7 edits across 1 file):
- Refreshed §Problem to accurately describe the tool-group-preserving
truncation helper and added ~5K fixed-overhead from system prompt
+ 19 tool definitions to the token-budget math
- Removed Story 5.1 docs-sweep deferral rationale (shipped in PR #60)
- Locked the JSONB-vs-table fork in §Proposed scope
- Added tool-call group invariant requirement +
chat_history_summarization_failed WARN fallback
- New §Open questions for /spec-gen with recommended defaults
- New §CLAUDE.md rule touchpoints (Rules #3, #5, #8, #10)
- Refreshed §Related work
* bug_fix.md — Investigation-mode /bug-fix output (149 lines):
- Problem / Reproduction / Root cause filled in with file:line
citations against agent_chat.py
- Owning layer locked: service; fix is additive (wrap existing
helper with summarization, don't replace)
- Fix design / Regression test / Rollout TBD pending user calls
on the 3 open forks
* MVP1_DASHBOARD.md + mvp1_dashboard.html — regenerated by the
mvp1-dashboard-regen pre-commit hook to reflect the new bug_fix.md
sibling (41 features total).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Code Review
This pull request documents the investigation and design for resolving chat history truncation issues. It introduces a detailed bug fix investigation report and updates the feature idea with technical specifications for implementing rolling summarization in MVP2. Feedback indicates that the auto-generated dashboard files contain truncated descriptions that break markdown links and HTML formatting, necessitating improvements to the generation script.
| | [chore_test_both_engines](../02_product/planned_features/chore_test_both_engines/idea.md) | Chore | `backend/tests/integration/test_clusters_api.py` only registers an **Elasticsearch** cluster in every test: | — | Idea (deferred from `infra_adapter_elastic` — refactor sweep, 2026-05-09) | | ||
| | [chore_trial_summary_single_query](../02_product/planned_features/chore_trial_summary_single_query/idea.md) | Chore | [`backend/app/db/repo/trial.py:aggregate_trials_summary`](../../../../backend/app/db/repo/trial.py) currently issues two SQL statements: | — | — | | ||
| | [bug_chat_long_conversation_truncation](../02_product/planned_features/bug_chat_long_conversation_truncation/idea.md) | Bug | `backend/app/services/agent_chat.send_user_message` defensively caps the OpenAI history at the most recent `HISTORY_MAX_MESSAGES = 100` messages and emits a `chat_history_truncated` WARN structlog lin | — | — | | ||
| | [bug_chat_long_conversation_truncation](../02_product/planned_features/bug_chat_long_conversation_truncation/idea.md) | Bug | [`backend/app/services/agent_chat.send_user_message`](../../../../backend/app/services/agent_chat.py) defensively caps the OpenAI history at the most recent `HISTORY_MAX_MESSAGES = 100` messages ([age | — | — | |
There was a problem hiding this comment.
The one-liner for this bug appears to be truncated, which breaks the markdown link and leaves the description incomplete. Since this file is auto-generated by the mvp1-dashboard-regen pre-commit hook, the generation script likely needs to be adjusted to handle long lines more gracefully, for example by truncating at a word boundary without breaking markdown syntax.
|
|
||
| </div> | ||
| <div class="one-liner">`backend/app/services/agent_chat.send_user_message` defensively caps the OpenAI history at the most recent `HISTORY_MAX_MESSAGES = 100` messages and emits a `chat_history_truncated` WARN structlog lin</div> | ||
| <div class="one-liner">[`backend/app/services/agent_chat.send_user_message`](../../../../backend/app/services/agent_chat.py) defensively caps the OpenAI history at the most recent `HISTORY_MAX_MESSAGES = 100` messages ([age</div> |
There was a problem hiding this comment.
The content for this one-liner div is truncated, which makes the description incomplete and includes broken markdown syntax. As this is an auto-generated file, the script that creates it should be updated to prevent truncating content in a way that breaks formatting and readability.
…s mid-markdown Idea file for the pre-existing bug in scripts/build_mvp1_dashboard.py that Gemini surfaced via F3 + F4 on PR #73. _extract_idea_problem caps prose at 240 chars via raw `para[:237] + "..."` with no awareness of markdown link balance, inline-code spans, or word boundaries. Includes regenerated MVP1_DASHBOARD.md + mvp1_dashboard.html (42 features total now that this folder is added). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Owner
Author
Gemini Code Assist review — adjudication
Net: 2 findings → 0 accept + 0 reject + 2 defer (both captured in Not blocking this PR's merge — the truncation behavior is pre-existing on main; this PR didn't introduce it. 🤖 Generated with Claude Code |
Fold the chore_mvp1_dashboard_truncation idea into this PR per the calibration discussion: ~30-LOC bounded fix + 13 unit tests is small enough to land inline rather than defer behind an idea file. Root cause: `_extract_idea_problem` was capping prose at 240 chars via raw `para[:237] + "..."` with no awareness of markdown link / inline-code / word boundaries. Fix: two new helpers — `_safe_truncate_markdown(text, max_len)` and `_strip_unclosed_markdown(text)` — replace the raw character cut with sentence-boundary preference + word-boundary fallback + strip unclosed [/]/(/)/backtick markdown + single-char ellipsis `…`. Tests: 13 cases in backend/tests/unit/scripts/test_dashboard_truncation.py (all pass locally). Regenerated MVP1_DASHBOARD.md + mvp1_dashboard.html with the new truncator. Deletes chore_mvp1_dashboard_truncation/ since the fix is no longer deferred. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
SoundMindsAI
added a commit
that referenced
this pull request
May 13, 2026
…coveries"
The existing rule ("either fix it now if inline-cheap or capture an
idea file") has been operating without a sharp definition of
"inline-cheap." Result: default lean toward capture, even for
medium-sized fixes that would have been faster to inline.
Adds a rubric table covering five discovery shapes — ≤20 LOC in same
file, ≤50 LOC scope-compatible, design-surface, CI-gate-protecting,
cross-subsystem — each mapped to one of four actions: inline / same-
branch adjacent commit / adjacent PR off main / idea file.
Plus a default-lean clarification: when borderline, lean inline. The
original rule was written when the failure mode was forgetting to
capture; this addresses the opposite failure mode.
Origin: dogfood discussion during PR #71/#72/#73 around the
chore_mvp1_dashboard_truncation discovery — captured as an idea
file initially, then user pushed back asking whether inlining would
have been faster. Folded the fix into PR #73 and codifying the
lesson here.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
4 tasks
Owner
Author
Updated adjudication — F3 + F4 now resolved by inline fixThe earlier verdict on F3/F4 was Defer (capture as
Tests: 13 unit cases in The 🤖 Generated with Claude Code |
SoundMindsAI
added a commit
that referenced
this pull request
May 13, 2026
* docs: add inline-fix vs idea-file rubric to CLAUDE.md "Tangential discoveries"
The existing rule ("either fix it now if inline-cheap or capture an
idea file") has been operating without a sharp definition of
"inline-cheap." Result: default lean toward capture, even for
medium-sized fixes that would have been faster to inline.
Adds a rubric table covering five discovery shapes — ≤20 LOC in same
file, ≤50 LOC scope-compatible, design-surface, CI-gate-protecting,
cross-subsystem — each mapped to one of four actions: inline / same-
branch adjacent commit / adjacent PR off main / idea file.
Plus a default-lean clarification: when borderline, lean inline. The
original rule was written when the failure mode was forgetting to
capture; this addresses the opposite failure mode.
Origin: dogfood discussion during PR #71/#72/#73 around the
chore_mvp1_dashboard_truncation discovery — captured as an idea
file initially, then user pushed back asking whether inlining would
have been faster. Folded the fix into PR #73 and codifying the
lesson here.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: tighten inline-fix rubric per Gemini F1+F2 on PR #74
F1 Accept: row 1 "no new tests needed" could be misread as a
shortcut to skip the Bug Fix Protocol's regression-test requirement.
Add explicit "NOT a bug fix; bug fixes always need a regression test"
carve-out so agents don't game the rubric.
F2 Accept: row 2 said ≤50 LOC, row 3 said >100 LOC — a 75 LOC
change fell into an undefined zone. Extend row 2 to ≤100 LOC so
the thresholds are continuous. (The principle is "bounded enough
that capture+resume costs exceed inline cost" — 100 LOC is
comfortably in that range for most scope-compatible fixes.)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
SoundMindsAI
added a commit
that referenced
this pull request
May 13, 2026
1. Dashboard KPI label "Features done" → "Scoped items done" in scripts/build_mvp1_dashboard.py (HTML and Markdown outputs). Sub-text explains the feat_/infra_/chore_/epic_ scope. Gemini flagged this on PR #76. Regenerated dashboards now show 17/17 (was 14/14 — incidental drift from Wave 1 archives). 2. Smoke test backend/tests/smoke/test_tutorial_path.py judgment- list deadline bumped 120s → 240s after two flake hits in one dev session (PRs #73 + #78, both passed on re-run). Doubled headroom, still well within the smoke job's 15-minute total. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
5 tasks
SoundMindsAI
added a commit
that referenced
this pull request
May 13, 2026
* chore: CI hygiene sweep — dashboard KPI label + smoke timeout 1. Dashboard KPI label "Features done" → "Scoped items done" in scripts/build_mvp1_dashboard.py (HTML and Markdown outputs). Sub-text explains the feat_/infra_/chore_/epic_ scope. Gemini flagged this on PR #76. Regenerated dashboards now show 17/17 (was 14/14 — incidental drift from Wave 1 archives). 2. Smoke test backend/tests/smoke/test_tutorial_path.py judgment- list deadline bumped 120s → 240s after two flake hits in one dev session (PRs #73 + #78, both passed on re-run). Doubled headroom, still well within the smoke job's 15-minute total. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(smoke): use time.monotonic for deadline; accept Gemini PR #79 Wall-clock time.time() could shift mid-test if NTP adjusts the system clock between deadline calculation and the while-loop check, causing early termination or extended hang. monotonic is the right primitive for duration measurement. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3 tasks
SoundMindsAI
added a commit
that referenced
this pull request
May 13, 2026
…skill (#81) Records the 2026-05-13 backlog sweep + new skill in state.md per the standard "update state.md whenever a feature lands or a priority shifts" rule: - Updates "Last updated" line to reflect the sweep - Adds 4 entries to "Most recent meaningful changes": - the 9-item backlog sweep (PRs #75-#80) - the /bug-fix skill (PRs #71-#72) + Investigation-mode dogfood on bug_chat_long_conversation_truncation (PR #73) - the CLAUDE.md inline-fix vs idea-file rubric (PR #74) - the docs-only paths-ignore CI filter (PR #70) - Expands "Active feature" line with the remaining-backlog breakdown (6 inline + 3 /bug-fix + 7 needs-second-pass + 4 keep-deferred) Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
SoundMindsAI
added a commit
that referenced
this pull request
Jun 3, 2026
…HSA-qx2v-qp2m-jg93) (#430) * chore(security): pin CI actions/images by digest + fix postcss vuln (GHSA-qx2v-qp2m-jg93) Resolves OSSF Scorecard supply-chain findings from the public code-scanning surface (security/code-scanning). Vulnerabilities (#72): postcss < 8.5.10 (moderate XSS via unescaped </style> in CSS stringify) was pulled in transitively as next@16.2.6's bundled postcss@8.4.31. Add a pnpm `overrides` for `postcss@<8.5.10` -> ^8.5.15 so the whole tree (including Next's copy) resolves to 8.5.15; regenerate the lockfile. `pnpm build` + full vitest (1008) green. PinnedDependencies (~60 alerts): pin every GitHub Action `uses:` ref to its 40-char commit SHA with a trailing `# vX` comment (56 refs across all 5 workflows), pin the 4 pr.yml service-container images (postgres, redis, elasticsearch, opensearch) by manifest digest, and pin the Dockerfile base images (node:26-bookworm-slim x3, python:3.14-slim via an ARG PYTHON_DIGEST, ghcr.io/astral-sh/uv:0.5.7) by digest. Both Dockerfiles pass `buildx --check`. Dependabot already covers github-actions + docker weekly, so it keeps the `uses:` SHA pins and Dockerfile FROM digests fresh. Deliberately left (impractical to hash-pin, low value, not "images"): - npmCommand `npm install -g pnpm@9` in ui/Dockerfile (#64/#65) — global tool install; corepack swap is a build-behavior change out of scope here. - pipCommand `pip install -r website/requirements.txt` in deploy-docs.yml (#73) — docs-site only; needs a hash-locked requirements file (--require-hashes). - Workflow `services.*.image` digests aren't auto-bumped by Dependabot (github-actions ecosystem updates `uses:` only) — manual refresh. Tier-3 Scorecard findings (branch protection, code-review ratio, project age, fuzzing, OpenSSF badge, SAST) are intrinsic/intentional and untouched. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: SoundMindsAI <eric.starr@soundminds.ai> * chore(security): consolidate digest-pinned base images into single BASE_IMAGE ARG Adjudicate both Gemini Code Assist findings on PR #430 (both accepted): - Dockerfile (high): a separate ARG PYTHON_VERSION + ARG PYTHON_DIGEST is a silent footgun — `--build-arg PYTHON_VERSION=3.13` would still pull 3.14 because the digest wins at pull time. Collapse to one ARG BASE_IMAGE (tag+digest together) so overrides are unambiguous. - ui/Dockerfile (medium): the node digest was triplicated across the deps/ builder/runner FROM lines. Declare it once as a top-level ARG BASE_IMAGE and reference ${BASE_IMAGE} in each stage. Both pass `docker buildx build --check`. Dependabot's docker ecosystem updates the `ARG ...=image:tag@sha256` form, so the pins stay fresh. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: SoundMindsAI <eric.starr@soundminds.ai> * docs(state): record chore_scorecard_pin_deps_postcss merge (PR #430) Prepend the PR #430 one-liner to state.md "Last 5 merges" (drop the now-6th chore_template_library_expansion into the older-entries line) + branch/Last- updated refresh; full narrative appended to state_history.md. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: SoundMindsAI <eric.starr@soundminds.ai> --------- Signed-off-by: SoundMindsAI <eric.starr@soundminds.ai> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
SoundMindsAI
added a commit
that referenced
this pull request
Jun 5, 2026
…card) (#456) Resolves two OSSF Scorecard Pinned-Dependencies findings (code-scanning #89, #73): the deploy-docs and build-guides-freshness workflows ran `pip install -r website/requirements.txt`, which pins exact versions but not by sha256 hash. - Add website/requirements.lock — a fully hash-pinned (sha256) lock compiled from the curated website/requirements.txt via `uv pip compile --generate-hashes`, covering the full transitive tree. - Both workflows now `pip install --require-hashes -r website/requirements.lock` (and key the pip cache on the lock). - website/requirements.txt stays the human-curated top-level source with a regenerate-the-lock pointer in its header; the lock carries a GENERATED / do-not-edit-by-hand header with the same regen command. requirements.txt is hand-bumped (Dependabot's pip ecosystem only watches the repo root, not website/), so there is no auto-update drift between the two files. Verified the require-hashes install succeeds in a clean venv and resolves mkdocs-material 9.7.6 / pymdown-extensions 10.21.3 / mkdocs-glightbox 0.5.2. Signed-off-by: SoundMindsAI <eric.starr@soundminds.ai> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Two work products landed together:
/idea-preflight+/bug-fix(Investigation mode) onbug_chat_long_conversation_truncation/.Files
idea.md(M)/idea-preflightAudit & Patchbug_fix.md(A)/bug-fixInvestigation modescripts/build_mvp1_dashboard.py(M)backend/tests/unit/scripts/test_dashboard_truncation.py(A) +__init__.py(A)MVP1_DASHBOARD.md+mvp1_dashboard.html(M)chore_mvp1_dashboard_truncation/idea.md(D)Dashboard fix detail
Root cause:
_extract_idea_problematscripts/build_mvp1_dashboard.py:127-139was capping prose at 240 chars via rawpara[:237] + "..."with no awareness of markdown link / inline-code / word boundaries. Most cuts landed in plain prose; after/idea-preflightonbug_chat_long_conversation_truncationthe cut shifted to land mid-[label](url)and broke the markdown.Fix: two new helpers —
_safe_truncate_markdown(text, max_len)and_strip_unclosed_markdown(text)— replace the raw character cut with: sentence-boundary preference (.,!,?within the last 50 chars) → word-boundary fallback (last space) → strip unclosed[/]/(/)/`` markdown → append single-char ellipsis…`.Before / after on the visible row:
Why expanded scope instead of separate PR
Conversation calibration: the original capture was a
chore_mvp1_dashboard_truncationidea file. The user pointed out we should lean toward inlining fixes when discovered if the work is bounded (~30 LOC + tests) rather than defer behind idea files. The rubric is being formalized in a separate CLAUDE.md PR.Test plan
pytest backend/tests/unit/scripts/)🤖 Generated with Claude Code