[Verification] /workflow real-machine scheduler validation — report implementation effects after local runs

[thiswind]

Purpose

Reminder issue: the /workflow command layer shipped in v0.20.0 (#6), but real-machine scheduler verification is still pending. After running local /workflow sessions, come back here and report the implementation effects.

What to verify (per carrier)

Carrier	Mechanism	What to confirm
Cursor	native background tasks / multitask	parallel dispatch, isolated sub-contexts, async result aggregation
Claude Code	run_in_background sub-agents	same as above
TRAE	task runner	same, plus serial-downgrade path for environments without a scheduler

Suggested test scenario

Reuse the recon scenario proven in the #8 dispatch experiment: one orchestrator, 3 read-only sub-agents on disjoint directories (e.g. _cursor/, _claude/, _trae_solo/), structured returns per SUBAGENT-DISPATCH.md.

Acceptance

• Cursor run: parallel OK, structured returns OK
• Claude Code run: parallel OK, structured returns OK
• TRAE run (or documented serial downgrade): OK
• Findings reported back to RFC: Decouple /workflow from Claude Code, cross-platform parallel workflow implementation based on native CAT architecture #6 — then RFC: Decouple /workflow from Claude Code, cross-platform parallel workflow implementation based on native CAT architecture #6 can be closed

References

• RFC: Decouple /workflow from Claude Code, cross-platform parallel workflow implementation based on native CAT architecture #6 — RFC (command layer shipped, verification pending)
• [RFC] Peer executors under CAT: /crew vs Claude Code Dynamic Workflows (/workflow or /ultra) #2 — original /workflow RFC (absorbed by RFC: Decouple /workflow from Claude Code, cross-platform parallel workflow implementation based on native CAT architecture #6, closed)
• Capability Test: Parallel Background Sub-Agent Execution #3 — cross-platform capability test report (absorbed, closed)
• SUBAGENT-DISPATCH.md — dispatch contract to follow during runs
• v0.20.0 CHANGELOG entry for the shipped command layer

[thiswind]

TRAE carrier: real-machine /workflow verification ✅ (2026-08-29)

Ran the suggested scenario in the live workshop session (PLAN-WF-002): one orchestrator (this session, frontier model) + 3 parallel read-only sub-agents dispatched via the TRAE task runner, per SUBAGENT-DISPATCH.md contracts.

Dispatch design (file-boundary split, read-only)

Sub-agent	Scope	Task
A. wf2_docs_mapping	AGENTS-GUIDE.md, SUBAGENT-DISPATCH.md, /workflow command body	verify docs match the shipped feature set
B. wf2_scripts_audit	_scripts/ (21 scripts)	inventory, dependency graph, dead-code suspects
C. wf2_version_audit	VERSION, CHANGELOG.md, git tags	version-discipline audit

Results — 3/3 structured returns compliant

• Parallel OK: all 3 dispatched in a single batch, no serialization observed.
• Structured returns OK: each returned the 4-part contract (summary / files_changed / verification / leftovers); all outputs landed in disjoint scratchpad dirs (ai_workspace/scratchpad/wf2_docs_mapping|wf2_scripts_audit|wf2_version_audit/), no file-boundary violations.
• Trust-but-verify held: the orchestrator re-checked every claimed artifact on disk — subtask A's README-footer finding (still v0.20.0) was confirmed and shipped as v0.20.2 (89bd462, tag + Release published).

Findings

1. Real bug found by the run itself: README "Current version" footer lagged one release behind — fixed and released as v0.20.2 (this issue's verification run doubling as a regression catch).
2. Docs mapping: role table / diagram / Features / How-it-works all reflect v0.20.0+; AGENTS-GUIDE persona map correctly Six masks.
3. Scripts audit: 21 scripts inventoried; suspects flagged for later triage (cleanup scripts write logs even in --dry-run; cleanup_topic_tree_temp.py possibly dead).
4. Version audit: 15 pre-0.15 tag/changelog mismatches (historical, not blocking); v0.19→v0.20.2 clean.

Status

• TRAE run: parallel OK, structured returns OK
• Cursor run — pending
• Claude Code run — pending

Remaining carriers documented here; #6 stays open until those two legs land. Full artifacts: workshop ai_workspace/scratchpad/wf2_*/.

[thiswind]

TRAE carrier: verification data available (structured returns, parallel dispatch)

Real-machine runs on the TRAE carrier (this machine, TRAE agent sessions using the task-runner mechanism):

Date	Run	Sub-agents	Result
2026-08-29	#8 dispatch experiment (3-way recon: persona map / scripts / workspace)	3 parallel, read-only, disjoint dirs	3/3 first-return compliance, structured returns (summary/facts/open_questions/leftovers) — archived in workshop ai_workspace/scratchpad/analysis/dispatch_log_20260829.md
2026-09-07	Tier-2 design recon (4-way: concurrency surfaces / installer-ignore policy / deployment-doc materials / upgrade-tool premises)	4 parallel, read-only	4/4 structured returns; findings directly fed five design docs (RFC-CONCURRENCY-001 etc.)

TRAE checkbox can be checked: parallel dispatch ✅ isolated sub-contexts ✅ structured returns ✅ (serial fallback not exercised — task runner available).

Plan for the remaining two carriers (merged into the v0.24.0 implementation round)

Per owner's coordination ruling tonight, the v0.24.0 implementation round itself will serve as the verification vehicle (dogfooding): its scope (5 guard scripts + 3 skills + cat_write gateway + validator refactor) splits cleanly into file-boundary-disjoint subtasks — we will run it as /workflow-orchestrated sessions on Cursor and on Claude Code, so implementation completion doubles as scheduler verification. Bonus: once the writes-skill gateway lands mid-round, the orchestrator-write-only topic-tree discipline gets machine-checked by the new gateway during the very verification runs.

Reporting back here after those runs; #6 closes when all three carriers report.

[thiswind]

v0.24.0 shipped — dogfooding evidence update

The one-shot v0.24.0 round (guard scripts + write gateway + operation skills) was itself executed as a /workflow-style orchestration in a TRAE session: 2 recon sub-agents + sequential gateway writes, all structured. Additionally cat_write.py topic-round served as the production path for this release's own topic-tree round — first real gateway write, journaled.

Remaining for this issue: Cursor-carrier and Claude-Code-carrier /workflow verification runs (planned as day-to-day work on those platforms; the design intent stands: implementation doubles as verification). v0.24.0 adds the machine-checked guardrails (gateway + doctor + closing) that those verification runs will exercise.