| name | fastworkflow-failure-archaeology |
| description | Load this when you are about to investigate, revert, re-attempt, or "clean up" something in fastWorkflow and suspect a battle may already have been fought: symptoms like "this bug looks old", "why does this dead code / legacy file exist", "should we just delete X", "the revert didn't revert", "bd close didn't stick", "payload is None in agent mode", "why was feature Y removed", or "did we already try this approach". Also load before proposing any feature that resembles: command dependency graphs, a standalone FastAPI service, probabilistic response generation, rich CLI, ensemble intent voting, or RLM-style memory layers — all have prior art documented here. Do NOT load for live triage of a current failure (use fastworkflow-debugging-playbook), for the rules governing changes (fastworkflow-change-control), or for how the architecture is supposed to work (fastworkflow-architecture-contract).
|
fastWorkflow Failure Archaeology
TEAM-PRIVATE: embeds content from uncommitted internal docs. Do not commit or publish this skill without the developer'sexplicit approval.
The chronicle of every major investigation, dead end, revert, and abandoned feature in this
repo — so no one re-fights settled battles. Each entry: symptom → root cause → evidence →
status. Every commit hash, path, and line number below was verified against the working tree
at v2.22.2 (commit c33b9a5) on 2026-07-09.
Prime directive of this skill: before you delete "dead code", revive an "obvious" feature,
or diagnose a "new" bug, check the index below. A spec doc in docs/ is NOT proof a feature
exists; an open beads issue is NOT proof work remains (tracker drift is real, see entry 11);
and a commit titled "Revert" is NOT proof anything was reverted (entry 4).
When to use / when NOT to use
| Situation | Use |
|---|
| "Has this been tried/broken/fixed before?" — history lookup | This skill |
| A test/runtime failure is happening NOW and you need triage steps | fastworkflow-debugging-playbook |
| You want the non-negotiable rules and the incidents behind them, to gate a change | fastworkflow-change-control |
| You need the intended design and its invariants | fastworkflow-architecture-contract |
| tau-bench / tau2 mechanics, pass@1 vs pass^k, our benchmark claims | fastworkflow-taubench-reference |
| Running the E0–E25 reliability campaign | tau2-reliability-campaign |
| Statistics for judging whether a "fix" is real (variance, pass^k math) | fastworkflow-proof-and-analysis-toolkit |
| Open research problems built on these failures | fastworkflow-research-frontier |
Index
| # | Entry | Status |
|---|
| 1 | The payload-loss bug (xray /invoke payload=None) | settled (v2.21/v2.21.1) |
| 2 | The TurnResult saga: what shipped (v2.21 X6 slice) vs v3.0 roadmap | partly shipped; v3.0 open |
| 3 | fix-85g: the 504-retry double-execution race (v2.22.0 turns engine) | Step 1 settled; Steps 2–3 open |
| 4 | The command_router.py revert trio (the revert that didn't) | moot since v2.6.0 |
| 5 | fix-0hb: the test suite that ate its own trained model | settled (fa97b48) |
| 6 | The RocksDB/speedict removal arc (v2.21.3–v2.21.6) | hot path settled; fix-7kp open |
| 7 | The fingerprint cache-invalidation bug (v2.22.1) | settled (b5747df) |
| 8 | The beads flakiness incident (2026-06-11) | workaround discipline, permanent |
| 9 | The private-doc auto-push incident (2026-07-08) | rule established, permanent |
| 10 | The RLM/tau2 failure taxonomy (Articles 1–3): the variance lesson | open — drives the tau2 campaign |
| 11 | Tracker drift: open beads issues describing shipped work | live caveat |
| 12 | The graveyard: abandoned features and dead code | reference table |
| 13 | Doc-rot register | live caveat |
Jargon used throughout, defined once:
bd / beads = the git-backed issue tracker (.beads/issues.jsonl is source of truth).
WEC = WorkflowExecutionContext, the transport-free execution core (Topology B).
CME = the internal command_metadata_extraction workflow that runs intent detection +
parameter extraction. DSPy ReAct = the LLM agent loop whose tools return text only.
RLM = Recursive/Reasoning Language Model — an LLM-written-code memory-retrieval layer used
in the team's tau2 experiments. tau2 (τ²-bench) = the retail/airline conversational-agent
benchmark; pass^k = task passes k independent runs in a row.
1. The payload-loss bug (xray /invoke payload=None)
- Symptom: xray's
POST /invoke returned payload=None, payload_hint="text",
request_id=-1 for every agent-mode data question, even when a table/chart was produced.
Deterministic /-prefixed commands returned the payload fine — that divergence was the tell
(docs/turn_result_design.md §0–§1.3).
- Root cause: NOT the mapping layer. Two framework drop points:
- The DSPy ReAct tool boundary is text-only.
_execute_workflow_query
(fastworkflow/workflow_agent.py) gets the full CommandOutput back from
CommandExecutor.invoke_command but joins only the .response strings and discards
artifacts — by design, because ReAct tools return string observations.
_finalize_agent_output synthesized a fresh answer-only CommandOutput from the
agent's final_answer, so every tool call's artifacts vanished from the user-facing path.
- Evidence:
docs/turn_result_design.md lines ~18–129 (origin narrative, code excerpts,
"confirming signature"); fix commits afcbe01 (v2.21.0) and aa20c1b (v2.21.1).
- Status: settled. Fixed at the chokepoint both transports flow through:
_finalize_agent_output now merges every artifact-bearing tool response into the single
answer CommandResponse.artifacts via merge_artifact_responses_into
(fastworkflow/workflow_execution_context.py, call at ~line 789; helper in
fastworkflow/turn.py). Key-collision rule: incoming key unchanged when absent, suffixed
_<n> on collision. Wire shape kept identical to the v2.20 baseline (one-element
command_responses).
- Near-miss worth remembering: v2.21 as first staged only captured turns into a dormant
TurnResult and left _finalize_agent_output unchanged — the bug would have shipped
"fixed" while still user-visible. Caught in a post-implementation teaching session
(2026-06-12, docs/turn_result_design_feedback.md Topic 5: "Don't merge v2.21 believing the
payload bug is fixed"). Lesson: re-verify the shipped code against the symptom, not the spec.
2. The TurnResult saga: what shipped vs the v3.0 roadmap
The payload bug triggered the largest design effort in repo history. Full detail:
references/turn-result-saga.md. The short version:
- Process: design doc (
docs/turn_result_design.md, with a decision log and amendments
A1–A47) → adversarial 48-finding review (docs/turn_result_design_review.md, R1–R48; beads
epic fix-vof with one child per finding, all 48 closed; one docs commit per resolution, from
42e9b3a "resolve R1" to 920d0de "epic fix-vof complete (48/48)") → ten-concern architecture
review (9031942, findings X1–X12) → authoritative spec docs/turn_result_design_final.md
("where this conflicts with any of those, this document wins").
- The X6 slice — what v2.21 actually shipped: the architecture review's X6 shrank the plan
from ~2.5k lines / ~120 files (96 command-file migrations) to ~500 lines / ~6 files
(
docs/turn_result_architecture_review.md lines 85–95). Shipped in afcbe01 + aa20c1b +
5a63790: fastworkflow/turn.py (TurnResult/TurnOutput/TurnStatus, mint_turn_key, artifact
merge helpers), the WEC turn accumulator, process_turn() returning the slim public
TurnOutput (workflow_execution_context.py:484), process_message deprecated, artifact
merge at finalize. TurnOutput.success is computed = all(co.success) — because v2.20
hard-coded the synthesized agent answer to success=True, masking every tool failure.
- What is NOT implemented (v3.0 roadmap,
turn_result_design_final.md §14–15): no
stores/ package, no ConversationTurnStore/ArtifactBlobStore, no Redis keyspace/ZSET
indexes, no MetricsSink, no fastworkflow admin CLI, no retention tooling;
run_fastapi_mcp/conversation_store.py still exists (slated for deletion at v3.0);
TurnResult.conversation_id/ordinal are Optional and left None. The
command_responses → command_response collapse is deferred to a v3.0 big-bang. Do not
assume any of these exist because the spec describes them.
- Collateral found by the review, still open: bug fix-5fv — CLI hangs on agent-mode
ask_user because the clarification CommandOutput is enqueued without a trace sentinel
(predicted by code-reading during R43; marked "NEEDS RUNTIME VERIFICATION"; bd show fix-5fv).
3. fix-85g: the 504-retry double-execution race (v2.22.0)
- Symptom: retrying
/initialize after a 504 returned tokens with a silently-empty
startup_output; occasionally two executions raced on the same context.
- Root cause: long LLM work (~10 min observed) ran inside the HTTP request.
asyncio.wait_for timeout raised HTTPException(504), which unwound async with runtime.lock, releasing the lock while the orphaned executor thread kept mutating the WEC —
so a retry passed the lock.locked() 409 guard and started a second racing execution.
- Evidence: commit cf3eeae;
bd show fix-85g; design doc
docs/fastworkflow_turns_async_execution_design.md; invariants pinned in the module
docstring of fastworkflow/run_fastapi_mcp/turns.py.
- Status: Step 1 settled (TurnRegistry, wait-or-defer never wait-or-abort, per-channel
active-execution pointer — never
lock.locked() — as the single liveness/idempotency
source, persist-before-DONE, idempotency-key rejoin). Steps 2 (GET /turns polling, 429
backpressure, TTL eviction, trace replay) and 3 (durable distributed turn store) are OPEN
(fix-85g.9–.13). Note /invoke_agent_stream still contains a runtime.lock.locked() check
(run_fastapi_mcp/__main__.py, grep for it) — a survivor of the old pattern.
4. The command_router.py revert trio (the revert that didn't)
- Symptom: debug cruft on an external contributor's parameter-extraction branch (Apr
2025): a hardcoded RocksDB cache path
./examples/sample_workflow/___convo_info/-694349230.db,
speedict Rdict helper functions, a commented-out not_what_i_meant retry block, and a
broad wildcard-fallback try/except.
- Root cause: the contributor (Sanchit Satija, all three commits 2025-04-08) ping-ponged
while trying to strip it: d2e3387 "Revert command_router.py changes" removed 58 lines;
aa751e7 "Revert changes" added 1 blank line; 1386f88 "revert changes" re-added 31 lines of
the same cruft. Net effect: the cruft shipped to main in the PR #10 merge.
- Evidence:
git show f9e2227:fastworkflow/command_router.py still contains
from speedict import Rdict, the dead helpers, and the commented -694349230.db block
(verified 2026-07-09).
- Status: moot.
command_router.py was dissolved in the v2.6.0 consolidation (a6d0f43,
"Now we have just 2 files - command directory and command routing"). The cruft survived on
main for roughly a year until then. Lessons: (a) verify a revert by diffing the file, not by
reading commit titles; (b) the deeper speedict addiction this cruft hinted at took until
v2.21.3/v2.21.4 to excise from the hot path (entry 6, where the still-open fix-7kp awaits
owner adjudication).
5. fix-0hb: the test suite that ate its own trained model
- Symptom: 4 tests in
tests/test_fastapi_service.py failed with FileNotFoundError on
fastworkflow/examples/hello_world/___command_info/global/threshold.json — but only on the
run AFTER a green run (v2.21.5 run: 4 passed; v2.21.6 run: 4 failed).
- Root cause: the module-scoped
trained_hello_world fixture in
tests/test_train_modern_stack.py trained against the REAL bundled example and rmtree'd its
___command_info at setup AND teardown, destroying the pre-trained model the FastAPI tests
depend on. The fixture only activated when datasets + real API keys were present, so it
was invisible in keyless CI — a local-only, every-other-run poison.
- Evidence: commit fa97b48;
bd show fix-0hb (full forensic narrative, "ROOT CAUSE
CONFIRMED"); the fix at tests/test_train_modern_stack.py (copytree into
tmp_path_factory.mktemp(...), ignoring ___command_info/___workflow_contexts/
___convo_info/__pycache__).
- Status: settled. Suite verified 478 passed / 0 failed and idempotent. Standing rule
(cite fix-0hb / fa97b48): never let tests or experiments train into or wipe
fastworkflow/examples/*/___command_info — always train into a temp copy. Bonus lore
recorded in the issue: the BUNDLED fastworkflow/examples/fastworkflow.passwords.env has a
dead placeholder Mistral key (litellm AuthenticationError); the repo-local
./passwords/.env has valid keys; local recipe:
fastworkflow train ./fastworkflow/examples/hello_world ./env/.env ./passwords/.env.
6. The RocksDB/speedict removal arc (v2.21.3–v2.21.6)
Three fixes in rapid succession, two of them on the same day (2026-06-15):
- v2.21.3 (79e6986) — env read at import time.
run_fastapi_mcp crashed at startup when
SPEEDDICT_FOLDERNAME lived only in --env_file: a module-level
session_manager = ChannelSessionManager() constructed the SessionStateStore in __init__
at import, BEFORE fastworkflow.init() loaded the env file. The bug was masked in-repo
because litellm calls load_dotenv on the cwd .env. Fix: lazy store construction.
Standing rule: never read env vars at module import time.
- v2.21.4 (6880b0a) — hot-path removal.
Workflow.create paid ~6 RocksDB Rdict
open/close cycles per new session (~90–270 ms of pure churn) for trivial payloads. Replaced
with a process-global dict + threading.RLock. Accepted regression, stated in the commit
body: "the CLI no longer resumes workflow context across process restarts." Durable state
is now exclusively SessionStateStore + ConversationStore (FastAPI only).
- v2.21.5 (033132f, fix-04r) — the leak the fix introduced, ~90 minutes later. Embedders
never call
close(), and a unique channel_id per request accumulated entries unboundedly.
Fix: the registry became a weakref.WeakValueDictionary (fastworkflow/workflow.py,
comment block at lines 14–40 tells the whole story) — GC auto-evicts abandoned sessions;
holding a strong reference IS the session.
- v2.21.6 (3fabdb0): security dependency updates rounding out the patch train.
- Status: hot path settled. speedict (
from speedict import Rdict) is still used,
intentionally, at: cache_matching.py, the enablecache decorator in workflow.py, the
NLU clarification cache (_workflows/command_metadata_extraction/intent_detection.py), and
run_fastapi_mcp/conversation_store.py; session_state_store.py stores plain JSON but
still keys its disk folder off SPEEDDICT_FOLDERNAME. Open-issue note — verified facts
only: fix-7kp (created 2026-06-16 00:32:40Z, OPEN) and fix-2jo (00:32:49Z, nine
seconds later, closed when v2.21.4 shipped) have byte-identical titles ("Remove
speedict/RocksDB from the Workflow hot path (Blueprint/Session-State split)") and
near-identical descriptions — both say "Ship as v2.21.4" and "speedict KEPT elsewhere".
Whether fix-7kp is a duplicate from the flaky-bd era (entry 8) or an intentional extension
is UNADJUDICATED — that call is reserved for the owner (see the stale-epic table in
fastworkflow-change-control): Ask Dhar; do not close fix-7kp yourself.
7. The fingerprint cache-invalidation bug (v2.22.1)
- Symptom: after adding/removing/renaming a command, runtime import errors pointed at
command sources that no longer existed, until the user manually deleted
___command_info
JSON artifacts.
- Root cause:
routing_definition.json had NO staleness check at all, and
command_directory.json used a newest-mtime check — which is blind to deletions and
renames (nothing bumps a surviving file's mtime when another file disappears).
- Evidence: commit b5747df;
compute_commands_source_fingerprint() at
fastworkflow/command_directory.py:627 (sha256 over the SET of (path, size, mtime_ns) of
_commands/**/*.py incl. base workflows + context model JSONs), stamped into both
artifacts; fastworkflow/command_routing.py rebuild-on-mismatch; orphaned
*_param_labeled.json pruned only AFTER a successful train so a failed retrain leaves the
prior artifacts runnable (fastworkflow/train/__main__.py).
- Status: settled. Doctrine: invalidate caches by content fingerprint of the full source
set, never by newest-mtime. Pre-2.22.1 workflows may carry unstamped artifacts (missing
fingerprint forces rebuild).
8. The beads flakiness incident (2026-06-11)
- Symptom:
bd close --reason=... reported success but silently failed to persist to
.beads/issues.jsonl (bare bd close worked). Separately, the Dolt server wedged into a
state where every bd command re-imported issues.jsonl into an empty database and async
exports raced/clobbered each other.
- Evidence: bd memory
beads-flakiness-observed-2026-06-11-in-the (read via
bd memories --json — note bd memories show <name> is not a valid invocation).
- Status: permanent working discipline. (1) One bd write at a time; (2) verify
.beads/issues.jsonl after each write; (3) never trust bd close --reason — use bare
bd close; (4) when the Dolt server wedges, patching issues.jsonl directly is safe because
every command re-imports it as source of truth. Possible collateral: the fix-2jo/fix-7kp
identical-title pair (entry 6 — duplicate vs intentional extension unadjudicated).
9. The private-doc auto-push incident (2026-07-08)
- Symptom: a private planning document was auto-committed and pushed to the PUBLIC
fastworkflow repo by an agent following a "session-close protocol". Cleanup required a git
history rewrite.
- Evidence: bd memory
never-git-commit-or-push-anything-without-dhar (established
2026-07-08).
- Status: permanent rule. NEVER
git commit or git push without the developer'sexplicit request
in that turn; no session-close protocol, skill, or workflow overrides this. Documents,
plans, and analyses may be team-private even when written into the repo tree — the repo is
public. This skill library itself is covered: it stays uncommitted until Dhar decides.
10. The RLM/tau2 failure taxonomy: the variance lesson
Source: three team-private PDFs in docs/ ("Article 1 - The Setup", "Article 2 - The Failure
Taxonomy", "Article 3 - Mitigations and a Path Forward", Jul 2026, principally Rakshit Pandhi
with Sanchit Satija and Sharvil Oza; all roles gpt-oss-20b on Groq; raw traces at
https://github.com/Programiz-007/fastworkflow_RLM_Traces/). System: an RLM memory layer for
τ²-bench retail — tool outputs over a token threshold τ are archived; retrieval is LLM-written
Python ending in SUBMIT().
- Headline numbers, on 15 hand-selected (enriched, non-random) retail tasks: pre-memory
~0/15 (asserted, never measured); initial memory-layer evaluation 2/15; after four bundled
mitigations 5/15. The variance lesson (Article 3, "Reproducibility"): the two tasks that
passed initially "were not among the five tasks that passed after the mitigations" — the
pass sets are DISJOINT. With n=15, single runs, and unfixed temperature, 2→5 is
statistically indistinguishable from noise. This is the founding argument for treating
pass^k reliability engineering as its own discipline and for E0 (harness + k=5 paired runs
- McNemar) being a hard prerequisite in the tau2 plan.
- The three failure pillars (Article 2):
- Intrinsic RLM failures — printed the answer instead of calling
SUBMIT, hitting the
20-iteration cap; repeated "not found" loops when the data was demonstrably present
(task with order #W1994898: the RLM hallucinated a 3-key index format and looped).
- Cross-layer failures — planner set
new_item_id equal to the original item
(9494281769→9494281769, a self-exchange the benchmark rejects); incomplete execution
reported as success ("illusion of success"); goal substitution (full cancellation
instead of modify_pending_order_items — the one bucket that got WORSE after
mitigations).
- Schema-awareness issues — queried a non-existent
product_type_id instead of
product_id.
Headline quote: moving retrieval into generated code "doesn't remove errors, it relocates
them" — into silently-wrong code that runs cleanly and SUBMITs a wrong answer.
- Known defects in the study itself (per the critical review folded into
docs/tau2_retail_reliability_implementation_plan.md §2.4): attribution was single-rater,
unblinded, forced single-label, and the counts don't sum (14 attributions vs 13 failures);
the four mitigations were bundled with no ablation; mitigation 4 (full tool-output view for
the planner) silently reintroduces the context bloat the memory layer exists to solve; the
self-exchange gold label is itself contestable. Twist verified against main: the current
retail example's exchange command already rejects self-exchanges (old∩new overlap check in
fastworkflow/examples/retail_workflow/_commands/exchange_delivered_order_items.py,
~lines 98–104) — implying the experiment fork bypassed validate_extracted_parameters or
ran stale wrappers. That plumbing audit is experiment card E1.
- Status: open. The whole tau2 retail reliability program (experiment cards E0–E25,
"Option N v2 / Governed Determinism") exists to turn this into measured science. See
tau2-reliability-campaign for execution and fastworkflow-taubench-reference for
benchmark mechanics. Also note the citation-hygiene sub-incident: Article 1 cites the
fastWorkflow paper DOI as 10.1145/3786335.3813158 (matching README.md:101) while Articles 2
and 3 cite 10.1145/3738331.3738402 — one is wrong (unresolved; verify externally before
citing either).
11. Tracker drift: open beads issues describing shipped work
process_turn()/TurnOutput are live on main (workflow_execution_context.py:484,
turn.py), yet epics fix-yy1 (v2.21 TurnResult capture) and fix-qtq (migrate the
run_fastapi_mcp transport edge to process_turn/TurnOutput) remain OPEN with open children.
This is partly deliberate ("NO COMMITS until user approval" epics whose approval/close loop
never ran) and partly real residue (run_fastapi_mcp/utils.py still has
run_process_message* helpers — the endpoint-level migration is genuinely incomplete).
Rule: reconcile beads state against git history before trusting bd ready or bd list --status open as a work queue. Stale cohort as of 2026-07-09: fix-5ka, fix-cgs, fix-4od
(all created 2026-06-04, untouched 35 days).
12. The graveyard: abandoned features and dead code
Do not resurrect, "finish", or silently delete anything here without checking the detail file:
references/graveyard.md. Summary:
| Item | Built | Killed / state | Why (or "unknown") |
|---|
| Command dependency graph (GenAI backward-chaining engine) | v2.10.0 (55a8668), refine built its JSON in v2.13.0 (ce5ada5) | tests removed v2.15.8 (ed21d77); sentencetransformers removed v2.17.19 (ad68d9b) | never adopted by the agent; superseded by injecting command info into the agent's system prompt. Orphan remnants still on main: fastworkflow/build/dependency_manager.py, fastworkflow/build/command_dependency_resolver.py (zero external callers) |
Standalone FastAPI service (services/run_fastapi, 1092-line main.py) | v2.16.0, 2025-10-11 (447db70) | replaced wholesale by fastworkflow/run_fastapi_mcp in v2.17.0, 2025-10-18 (299739e) | lived one week; v2.17 line absorbed 35 patch releases and became the product |
| Probabilistic response generation | branch origin/cursor/implement-probabilistic-response-generation-and-pass-tests-7586 (7c7d1cf, a4d3d4d; forked from ddab25c, Aug 2025) | never merged; fastworkflow/probabilistic_response.py does not exist on main | why-abandoned unknown. Trap: the spec docs/probabilistic_response_generation.md IS on main — a spec doc is not proof the feature exists |
| Rich CLI (replace colorama with rich) | branch cursor/enhance-fastworkflow-cli-aesthetics-1752 (4d2d9a7, c4cf167; forked from v2.5.2) | never merged; main still depends on colorama (pyproject.toml:49) | why-abandoned unknown |
| 43-agents cloud-agent experiment ("agent categorization system with 43 agents in 4 collections") | branches origin/cursor/new-cloud-agent-1d34 / -af26 (2a0214f etc., Feb 2026, forked from v2.17.33) | never merged; nothing on main | Cursor cloud-agent experiment; why-abandoned unknown |
majority_vote_predictions (5-way parallel intent voting) | present at fastworkflow/_workflows/command_metadata_extraction/intent_detection.py:304 | dead code: sole call site commented out (line 122) | its own TODOs (lines 302–303) admit generation is deterministic so all 5 votes return the same answer; needs temperature support first. Open question: revive or delete |
ambiguous_threshold.json legacy files | pre-split threshold artifact | linger in trained folders (e.g. fastworkflow/_workflows/command_metadata_extraction/___command_info/*/); the loader reads only tiny_ambiguous_threshold.json / large_ambiguous_threshold.json (model_pipeline_training.py:299-300) | harmless orphans from a rename; safe to ignore, not proof of a live code path |
13. Doc-rot register
Documented contradictions where code wins over docs (evidence attached; do not "fix" docs here
without change control):
- CLAUDE.md says intent models are trained "via scikit-learn". Actually torch/transformers
fine-tuning of a two-tier TinyBERT/DistilBERT pipeline; sklearn supplies only
LabelEncoder/split/metrics (see
fastworkflow/model_pipeline_training.py imports). CLAUDE.md
also omits tests/todo_list_workflow from the test-workflow list.
- README/Article DOI disagreement — see entry 10.
- A spec doc in
docs/ is not proof of implementation — probabilistic_response_generation.md
(entry 12) and the TurnResult v3.0 sections (entry 2) are the standing examples.
run_fastapi_mcp/README.md still references the deleted services.run_fastapi layout
(entry 12's one-week service).
Provenance and maintenance
Facts verified 2026-07-09 against v2.22.2 (commit c33b9a5). Team-private sources embedded:
the three Article PDFs, docs/tau2_retail_reliability_implementation_plan.md,
docs/rsi_harness_agent_report.md (all untracked in docs/). Re-verification one-liners for
every volatile fact:
| Fact | Re-verify with |
|---|
| Commit hashes/subjects cited anywhere above | git log --oneline --no-decorate -1 <hash> |
| Revert-trio cruft shipped to main | git show f9e2227:fastworkflow/command_router.py | grep -n "speedict|-694349230" |
| command_router.py dissolution | git log --follow --oneline -- fastworkflow/command_router.py | head -3 |
| Payload fix wired at finalize | grep -n merge_artifact_responses_into fastworkflow/workflow_execution_context.py |
| process_turn line number | grep -n "def process_turn" fastworkflow/workflow_execution_context.py |
| v3.0 pieces still absent | ls fastworkflow/stores fastworkflow/turn_serializer.py 2>&1; ls fastworkflow/run_fastapi_mcp/conversation_store.py |
| fix-85g / fix-7kp / fix-5fv / fix-yy1 / fix-qtq open status | bd show <id> --json (read-only) |
| fix-vof child count (48) | python3 -c "import json;print(sum(1 for l in open('.beads/issues.jsonl') if json.loads(l).get('id','').startswith('fix-vof.')))" |
| fix-2jo/fix-7kp identical titles + 9 s timestamps | grep -E '"id":"fix-(2jo|7kp)"' .beads/issues.jsonl | python3 -c "import sys,json;[print(json.loads(l)['id'],json.loads(l)['created_at'],json.loads(l)['title']) for l in sys.stdin]" |
| bd memories (flakiness, push rule) | bd memories --json |
| fix-0hb temp-copy fix in place | grep -n "copytree|tmp_path_factory" tests/test_train_modern_stack.py |
| speedict remaining sites | grep -rln speedict fastworkflow/ --include=*.py |
| Fingerprint function | grep -n "def compute_commands_source_fingerprint" fastworkflow/command_directory.py |
| workflow.py weakref registry + CLI-resume trade-off comment | sed -n '14,45p' fastworkflow/workflow.py |
| majority_vote dead code + TODOs | grep -n majority_vote fastworkflow/_workflows/command_metadata_extraction/intent_detection.py |
| ambiguous_threshold.json orphans | find fastworkflow -name ambiguous_threshold.json |
| Dep-graph orphan modules have no callers | grep -rn "dependency_manager|resolve_command_dependencies" fastworkflow/ --include=*.py | grep -v "build/dependency_manager.py|build/command_dependency_resolver.py" |
| Abandoned branches unmerged | git branch -r --merged main | grep cursor (expect empty) |
| Self-exchange overlap check on main | sed -n '98,104p' fastworkflow/examples/retail_workflow/_commands/exchange_delivered_order_items.py |
| Article facts (0→2→5, disjoint sets, pillars) | extract text: PYTHONPATH=/tmp/pdflib .venv/bin/python -c "import pypdf;print(''.join(p.extract_text() for p in pypdf.PdfReader('docs/Article 3 - Mitigations and a Path Forward.pdf').pages))" | grep -i "not among" (install pypdf to a temp target first) |
| DOI disagreement | grep 10.1145 in extracted Article texts vs grep -n 10.1145 README.md |
| Sanity-run the whole table | bash .claude/skills/fastworkflow-failure-archaeology/scripts/verify_archaeology.sh |
Maintenance protocol: when a new investigation closes, append an entry here in the same
symptom → root cause → evidence → status shape, with verified hashes, and add its
re-verification one-liner to the table. Never rewrite an existing entry's history — append a
dated correction (the amendments-over-edits discipline from the TurnResult saga).
