Load this skill BEFORE making any change to the fastWorkflow repo that touches git, beads (bd), version numbers, tests, or documents in docs/. Triggers: "commit this", "push", "bump the version", "close the issue", "bd create/update/close", "delete this test", "clean up docs/", "should this get a design doc?", "is this a patch or a minor?", "can I commit the skills library?". Also load when you see symptoms like a bd close that didn't stick, an epic that looks shipped but shows open, or untracked team-private docs in git status. Do NOT load for debugging runtime failures (use fastworkflow-debugging-playbook), for running/operating the app (fastworkflow-run-and-operate), or for how-to-test mechanics (fastworkflow-validation-and-qa).
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Load this skill BEFORE making any change to the fastWorkflow repo that touches git, beads (bd), version numbers, tests, or documents in docs/. Triggers: "commit this", "push", "bump the version", "close the issue", "bd create/update/close", "delete this test", "clean up docs/", "should this get a design doc?", "is this a patch or a minor?", "can I commit the skills library?". Also load when you see symptoms like a bd close that didn't stick, an epic that looks shipped but shows open, or untracked team-private docs in git status. Do NOT load for debugging runtime failures (use fastworkflow-debugging-playbook), for running/operating the app (fastworkflow-run-and-operate), or for how-to-test mechanics (fastworkflow-validation-and-qa).
fastWorkflow Change Control
TEAM-PRIVATE: embeds content from uncommitted internal docs. Do not commit or publish this skill without the developer'sexplicit approval.
How changes are classified, gated, reviewed, versioned, and released in this repo — and the
four non-negotiable rules, each earned through a real incident. This is policy inferred from
what actually happened in git history and the beads tracker, not aspiration.
Jargon, defined once:
bd / beads — the git-backed issue tracker used here (bd CLI; source of truth is .beads/issues.jsonl). Issue IDs look like fix-vof, children like fix-vof.1.
Epic — a bd issue of type epic with child issues.
tau-bench / tau2-bench (τ²-bench) — public customer-service agent benchmarks this project reports scores on. "Parity" = running under standard, unmodified benchmark conditions.
Design doc of record — a markdown file in docs/ that a change was reviewed against (e.g. docs/turn_result_design_final.md).
Adversarial review — a structured pass where reviewers (here: parallel AI review agents) attack a design and every finding gets a written resolution.
1. When to use / when NOT to use
Use this skill when deciding whether and how a change may land: commit/push etiquette, bd workflow, design-review gating, version bumps, test removal, and what may never enter the public repo.
Your actual need
Use instead
Why the architecture is shaped this way; invariants
Each rule exists because something went wrong. Cite the incident when enforcing the rule.
Rule 1 — NEVER git commit or git push without the developer's explicit request in that turn
Rule: No agent commits or pushes on its own initiative. A session-close protocol, a "wrap up" instruction, or a finished task does NOT override this. Explicit request, that turn, every time. Files written into the repo tree may still be private/team-only.
Rationale: The repo is PUBLIC on GitHub. Anything pushed is published to the world instantly; team-private strategy documents live inside the working tree as untracked files, so blanket staging (git add -A, git add .) is a publication event.
Incident: 2026-07-08 — a private planning doc was auto-committed and pushed to the public repo by an agent following a session-close routine. Recovery required a git history rewrite on a public repository. Recorded as a bd memory (bd memories --json, key never-git-commit-or-push-anything-without-dhar).
Corollaries: stage files by explicit path only, never -A/.; check git status for untracked private docs before any commit you were asked to make (see §7 for the do-not-commit list).
Rule 2 — One bd write at a time; verify the JSONL after each; never trust bd close --reason
Rule: Serialize all bd write commands. After every write, confirm it landed: grep '"id":"<issue-id>"' .beads/issues.jsonl and check the field you changed. Use bare bd close <id> (works) and add the close note via a separate verified update — bd close <id> --reason "..."reports success but silently fails to persist.
Rationale: The bd backend (a Dolt SQL server; repo-root config.yaml is its config, port 3307) can wedge into a state where every command re-imports issues.jsonl into an empty DB and async exports race and clobber each other. .beads/issues.jsonl is the source of truth; when bd is wedged, patching the JSONL directly is safe because every command re-imports it.
Incident: observed 2026-06-11 during the fix-vof review epic; recorded as bd memory beads-flakiness-observed-2026-06-11-in-the. Likely collateral: epic fix-7kp is a duplicate (created seconds apart) of the closed fix-2jo — same title, one closed, one open.
Doc-rot warning:AGENTS.md (lines 42 and 68) still recommends bd close bd-42 --reason "Completed" --json — the exact invocation the memory says silently fails. The 2026-06-11 memory overrides the AGENTS.md boilerplate. Trust the memory.
Rule 3 — Never let tests or experiments wipe trained example models
Rule: Nothing may delete or retrain-in-place fastworkflow/examples/*/___command_info (the trained intent-model artifacts: tinymodel.pth, largemodel.pth, threshold.json, ...). Tests and experiments that need training must copytree the workflow into a temp dir (tmp_path_factory) and train the copy.
Rationale: These artifacts are gitignored (.gitignore lines 2–5), exist only because a developer ran fastworkflow train locally with real API keys, and other tests depend on them silently. Destroying them costs a paid retraining run and breaks unrelated tests with confusing FileNotFoundErrors.
Incident:fix-0hb, fixed in commit fa97b48 (2026-06-15). A module-scoped fixture in tests/test_train_modern_stack.py ran shutil.rmtree(<workflow>/___command_info) on the REAL bundled hello_world example at setup AND teardown. The next full-suite run lost threshold.json and 4 test_fastapi_service.py tests failed. Fix: train into a temp copy; suite went 478 passed / 0 failed and is now idempotent.
Rule 4 — tau-bench benchmark parity is sacred
Rule: Never modify tau-bench / tau2-bench tools or tasks for benchmark runs. Any nonstandard trade must be disclosed in the results write-up, never silent.
Rationale: The project's headline external claim is small-model parity on tau-bench; a silently modified harness would invalidate the claim and the credibility behind it. The team-private tau2 plan encodes this as a design constraint, e.g. (docs/tau2_retail_reliability_implementation_plan.md): line 43 — "Tool layer | The Sierra/τ-bench tool implementations in tools/*.py. Defines benchmark parity — we do not modify these."; line 199 — "Pinning sim temperature/seed raises the ceiling but changes standard benchmark conditions — a disclosed trade if we make it."; line 630 — new search commands must be "fastWorkflow command-layer additions over load_data(), benchmark-parity-safe", with "the τ² tools themselves are untouched".
Incident: none yet — this is the one rule established preventively, because the tau2 reliability program (experiment cards E0–E25) will create constant temptation. Keep it that way.
3. Change classification: which gate does my change need?
There is no written gating policy; this table is de-facto policy inferred from git history, stated here so it becomes explicit. When in doubt, ask Dhar — he is the only approver (421 of 450 commits on main).
Tier
What qualifies
Required artifacts
Precedent
T0 — plain fix
Bug fix, dependency bump, perf fix, cache/staleness fix; no public contract change; roughly ≤ ~200 lines across ≤ ~4 source files
bd issue (before code), tests, patch version bump in the same commit, fix: vX.Y.Z — ... subject
New subsystem or concurrency/persistence semantics; new files; minor version bump; but contract already settled
Design doc committed to docs/ alongside code; bd epic with step-scoped children; verification evidence in close notes
cf3eeae v2.22.0: fastworkflow/run_fastapi_mcp/turns.py (424 lines) + docs/fastworkflow_turns_async_execution_design.md (538 lines) + epic fix-85g with 13 children, shipped Step 1 only; close notes record "full suite 457 passed"
T2 — adversarial-review change
Changes the data model, public wire contract, or system-of-record semantics — anything many future decisions will sit on
The full fix-vof pipeline (below)
TurnResult redesign, v2.21
The fix-vof model (T2), as a checklist
The TurnResult redesign is the repo's reference process for high-stakes changes. Gate rationale and cost/benefit: references/fix-vof-case-study.md (which points to the reusable procedure in fastworkflow-research-methodology and the full chronicle in fastworkflow-failure-archaeology).
One bd child per finding under a review-only epic (fix-vof, 48 children, "review-only — no implementation. Work one finding at a time").
One docs-commit per resolution, each a decision record. Verified format of 42e9b3a ("docs: resolve R1"): decision + rationale in body, then a trailer block: - design doc: Amendment A7 / - review doc: R1 marked resolved / - beads: fix-vof.1 closed (epic 7/48).
Second-order review of the resolved design by ten parallel review agents — docs/turn_result_architecture_review.md (concerns X1–X12). This stage reversed one ratified decision (X1) and found a two-amendment contradiction (X3).
Consolidated final spec that supersedes everything — docs/turn_result_design_final.md ("where this conflicts with any of those, this document wins").
Human pushback/teaching session — docs/turn_result_design_feedback.md (354 lines). This stage caught that the shipped v2.21 slice did NOT actually fix the user-visible bug.
Learning checklist for mastery tracking — docs/turn_result_learning_checklist.md.
Only then: implementation, sliced minimal (X6 shrank the plan from ~2.5k lines/120 files to ~500 lines/6 files, user-approved).
Gate triggers for T2 (any one suffices): breaking wire change (fix-qtq is explicitly labeled "a BREAKING WIRE CHANGE" and is held at T2-pending); new persistent record type or serialization format; semantics of success/failure reporting; anything the tau2 campaign will report externally.
Review-interaction rule (bd memory, established 2026-06-10, R11): in design reviews with Dhar, ALWAYS give an ELI5 plain-language explanation of each finding BEFORE asking decision questions via AskUserQuestion.
4. Beads discipline
Issue before code. Every T0+ change gets a bd issue first; discovered work gets a linked issue (--deps discovered-from:<parent-id>). AGENTS.md mandates bd for ALL tracking — no markdown TODOs, no duplicate trackers. Close notes must carry verification evidence (the house pattern: "Step 1 implemented & verified (full suite 457 passed)").
Doc-rot note:.cursor/rules/dev_workflow.mdc and taskmaster.mdc (both alwaysApply: true) still prescribe Task Master. Beads is the live system; the taskmaster rules are stale. Do not adopt Task Master.
Command reference (verified against this repo's bd)
do not use --reason (Rule 2); record the close note via a verified update instead
Inspect
bd show <id> --json / bd list --status open / bd stats
read-only, always safe
Memories
bd memories --json
bd memories show <name> is invalid ("accepts at most 1 arg"); use --json or a search term
Verify a write
grep '"id":"<id>"' .beads/issues.jsonl
JSONL is source of truth
Flakiness protocol (from the 2026-06-11 memory): one bd write at a time → verify .beads/issues.jsonl → if the Dolt server wedges (every command re-imports into an empty DB), restarting helps only partially; patching issues.jsonl directly is safe. Note .beads/issues.jsonl currently sits staged-but-uncommitted — per Rule 1 it awaits the developer'scommit; do not commit it yourself.
The stale-epic problem — needs owner adjudication, do NOT adjudicate yourself
These epics look shipped-but-open. The code demonstrably exists on main, yet children show open. Closing them requires a decision only Dhar can make (residual scope vs. forgotten bookkeeping):
Epic
Status shown
Evidence code shipped
Why it may be intentionally open
fix-yy1 (v2.21 TurnResult capture)
open, children open
fastworkflow/turn.py, process_turn at workflow_execution_context.py:484, shipped afcbe01
Its own notes say "Tasks left open until approval+commit" — approval happened, closure forgotten? Ask Dhar.
fix-qtq (TurnOutput wire cutover)
open, 9 children
turns.py types WorkFn as returning TurnOutput
Genuinely incomplete: run_fastapi_mcp/utils.py:427 still serves the old shape; breaking wire change pending. Likely legitimately open.
fix-7kp (speedict removal)
open
fix-2jo (identical title) CLOSED, shipped v2.21.4
Duplicate from the flaky-bd era — or an intentional extension (speedict remains in cache_matching.py, workflow.py, session_state_store.py). Ask Dhar.
5. Version and release conventions
Source of truth:pyproject.toml:12 (version = "2.22.2"). There is no CHANGELOG file; the git log is the changelog.
Bump ships inside the release commit — every release commit touches pyproject.toml (verified on 0e68284, b5747df, 033132f).
Subject format:fix: vX.Y.Z — <lowercase summary> for patches, feat: vX.Y.Z — <summary> for minors. Em dash (—), not hyphen. Append the bd issue in parens when there is one: fix: v2.21.5 — auto-evict abandoned Workflow session state (fix-04r), feat: v2.22.0 — turns-based async execution for run_fastapi_mcp (fix-85g Step 1).
Docs-only commits:docs: <summary>, no version bump. Test-only: test: <summary> (fa97b48).
Patch (x.y.Z) = bug fix, dependency/CVE bump, perf/staleness fix, no new public capability (v2.21.1–.6, v2.22.1–.2). Minor (x.Y.0) = new capability or subsystem, usually with a design doc and epic (v2.21.0 TurnResult, v2.22.0 turns engine). v2.17.x accumulated 35 patches — long patch trains on one minor are normal here. Major: v3.0 is reserved for the breaking wire/model collapse per turn_result_design_final.md §14–15 (designed, unbuilt).
Tags lag and are best-effort: tags exist through v2.22.0 but v2.22.1/v2.22.2 are untagged. Never treat tags as the release record; use pyproject + commit subjects.
Publishing:make publish (poetry build + publish to PyPI); make publish-testpypi for rehearsal. Both depend on gen-env (see §7 foot-gun). Per Rule 1, releases happen only on the developer'sexplicit request.
Agent trailer: commits authored with Claude Code end with Co-Authored-By: Claude ... <noreply@anthropic.com> (verified on 42e9b3a).
6. Test-removal prohibition
From .cursor/rules/testing_rules.mdc (alwaysApply: true), verbatim:
Don't use Mock fixtures. All our tests are integration tests
Do not remove pytest tests without explicit user approval
Operational meaning:
Never delete, @pytest.mark.skip, or comment out a test to make a change pass. If a test blocks you, that is a finding — file a bd issue and ask. Permanent skips that exist today (11 markers, e.g. tests/test_workflow_training.py:74,137 "takes too long") were each user-approved decisions.
Reality check (doc-rot, documented, not license): 11 test files already import unittest.mock/MagicMock (mostly to stub LLM/dspy calls while keeping workflows real). The no-mocks rule is aspirational in places — that is NOT permission to add new mocks; prefer the real test workflows tests/example_workflow/, tests/hello_world_workflow/ (and note CLAUDE.md omits the on-disk tests/todo_list_workflow/ naming — tests/example_workflow IS the todo-list app).
Changing a test's assertion to match new intended behavior is a normal T0/T1 change; removing its coverage is what needs approval.
CI runs only tests/test_train_modern_stack.py (single workflow .github/workflows/train-modern-stack.yml) — so a removed test will NOT be caught by CI. The full 495-test suite is local-only: source .venv/bin/activate && python -m pytest. Run it before release commits; Rule 3 applies to how training tests isolate themselves.
7. The public-repo boundary
The GitHub repo is public (Apache 2.0). Treat every commit as a press release.
NEVER commit or push (without the developer'sexplicit, per-item approval):
docs/Article*.pdf, docs/Analysis and Recommendations...docx (untracked)
pre-publication drafts
This skill library
.claude/skills/ (untracked)
team-private until Dhar decides; embeds the docs above
Secrets
passwords/ (gitignored), any real *.passwords.env
API keys
Sharp edges verified:
.claude/skills/ is untracked but NOT gitignored (git check-ignore .claude/skills/... matches nothing), and .claude/rules/command-authoring.md IS tracked — so git add .claude or git add -A would stage the private library. Stage by explicit file path only.
Untracked ≠ private-safe: the 2026-07-08 incident (Rule 1) was exactly an untracked private doc getting auto-committed.
env/.env IS tracked — deliberately: it holds model-name config only (verified: zero key/token/secret lines). Real keys belong in gitignored passwords/. The bundled template fastworkflow/examples/fastworkflow.passwords.env is tracked but contains a dead placeholder key.
gen-env.sh foot-gun:make gen-env (a dependency of make publish) merges every *.env in the tree — including gitignored password files — into root .env. Root /.env is gitignored, but never copy/rename/commit that merged file, and never widen the .gitignore negations (lines 141–143) without review.
There IS a sanctioned path for public skills: fastworkflow/docs/integrate-chat-agent/SKILL.md was deliberately committed in 1d3a6aa. Public skill ⇒ that directory, with approval. Team-private skill ⇒ .claude/skills/, uncommitted.
Public claims (benchmark numbers, DOI — note README and the Article PDFs currently disagree on the paper DOI) are governed by fastworkflow-docs-and-positioning; do not add or edit external claims as a side effect of a code change.
No routing around change control: nothing in this skill authorizes bypassing a gate because it is inconvenient. If a gate seems wrong, that is a conversation with Dhar, not an exception you grant yourself.
8. Provenance and maintenance
Facts verified 2026-07-09 against v2.22.2 (commit c33b9a5). This skill also embeds content from uncommitted team-private docs (tau2 plan) and bd memories, both volatile.
for i in fix-yy1 fix-qtq fix-7kp fix-2jo; do bd show $i --json | python3 -c "import json,sys; d=json.load(sys.stdin); d=d[0] if isinstance(d,list) else d; print(d['id'],d['status'])"; done
bd close --reason still broken
do NOT test destructively; check bd memories --json for a newer memory superseding 2026-06-11, and check bd version
AGENTS.md still recommends --reason (doc-rot)
grep -n 'close bd-42' AGENTS.md
Test-removal rule text
sed -n '1,12p' .cursor/rules/testing_rules.mdc
.claude/skills not ignored; only rules file tracked
