docs(WS-0): truth pass across six workstreams — the board was describing a codebase from weeks ago

[vjvarada]

Truth pass across six workstreams

Thirteen of thirteen board rows have now failed the seven-point contract, always on the same point: no testable done-when. The six that had never been audited were audited together — all six NO-GO.

The cause turned out not to be missing work. The specs describe a codebase from weeks ago, so they ask for things that already exist — and in one case would have caused an implementer to break working code.

🔴 The one that would have done damage

permissions_sandbox_b6.md asserts, as "grep-confirmed", that the mutation container runs with zero --cap-drop/--memory/--cpus/--pids-limit flags. Four of those six have been present since 2026-07-27. An implementer dispatched on WS-3 would most likely have re-added flags that already exist.

The same spec never recorded that Tier 0 shipped, that a containerized Copilot CLI runtime shipped and is wired at two call sites behind a flag that ships "", or that isolation_tier() has been computed and thrown away into a log field since it landed.

The pattern, row by row

Row	What the spec asked for vs. what is true
WS-1	"Remaining: Zoho handlers" is fiction — the Zoho client is read-only (six list_* calls, two GETs), with no write path anywhere to broker. Two real flip-blockers surfaced instead, neither previously written down
WS-3	Above. Plus "T2 for non-first-party agents" — no such field exists anywhere in the code; it is a heading with no data model
WS-8	manifest.py and declarative.py are built, tested and unwired — ~60% of Phases A+B is on disk. manifest.py's own docstring said "nothing here is wired into the run path yet": the single sentence most likely to cause a duplicate build
WS-11	Full-graph copilot authoring shipped as F14 — the same document said so twice. Parallel fan-out shipped too. The real content is fan-in and loops
WS-12	~90% delivered elsewhere. A high-severity risk had retired itself — openai is already on 2.x, so Phase 4 drags one SDK major, not two
WS-21	Two of four deliverables already shipped, and two of the three gtd_time_blocks acceptance clauses were satisfiable by doing nothing

Two findings that are live defects, not documentation

WS-1, and both block the ACTION_BROKER_ENFORCE flip:

1. The two irreversible ClickUp actions (delete_task, archive_task) are gated but have no registered handler — six action names gated, four handlers. Approving a queued delete hits the no-handler branch and marks the row failed. The existing test pins the set to a four-element literal and is structurally blind to it.
2. A broker-queued push is written as sync_state='synced' with an empty provider_task_id — items marked synced to nothing.

The kill-switch's own docstring now says not to flip it until both land.

Corrections that change sequencing

• "Phase A unblocks D3" is false in the direction that matters. config.json-based instancing already ships via AgentManifest.instance_key(), so Centers C was never waiting on WS-8 Phase A — the board has claimed otherwise for weeks.
• WS-11's two new tickets must each invert a currently-pinned test (test_fan_in_rejected_v1, test_cycle_rejected). Without that inversion they close green having built nothing — the same failure shape BO-20b was caught on.
• calendar_timeboxing.md's "non-breaking TimeBlock[] swap" claim was false. The real blast radius is 17 TS files + 3 gateway modules + a skill + an agent, so gtd_time_blocks is 4 PRs, not 1.

🔵 Owner decisions recorded as D10

1. Command Center is an internal Fracktal tool. WS-3's full run sandboxing (T2) is therefore parked under a trusted-colleague threat model — design intact, schedule removed, with an explicit un-parking condition (a second org, or agent authorship from outside Fracktal) rather than carried as debt.
2. Loops in the workflow engine are approved, against workflows_app.md §11's standing anti-n8n rule — which is clarified to govern the node catalog, not the control-flow vocabulary, so it can't be mis-cited as a blocker later. The engine-complexity cost is stated honestly.

What all six specs now have

Each leads with what is actually built, carries per-item done-whens a command or an assertion can settle, names verification commands by file (never tests/unit/ as a directory, which hangs against the live DB), and labels every item AGENT-SAFE or OWNER-GATE — there were zero such labels across all six before.

Board-side: four rows added to the single-owner registry, six to the owner-gate registry (including copilot_sandbox_scope, which routes real Copilot sessions into containers and was registered nowhere), one wrong gate anchor fixed, two §5 residuals closed and one opened. The Horizons ownership dispute between WS-21 and WS-18 is assigned rather than left to be rediscovered.

Verification

uv run ruff check . --select F821,F601,F602,F502,F7,B006     -> All checks passed!
uv run pytest tests/unit/test_action_broker.py tests/unit/test_actions_routes.py \
  tests/unit/test_provider_broker_gate.py tests/unit/test_task_broker_handlers.py \
  tests/unit/test_agent_manifest.py tests/unit/test_declarative_builder.py -q
                                                              -> 94 passed in 1.98s
23 WS rows, no duplicates, pipe counts uniform within every table block

Per-spec fences measured by their own agents and now recorded in the specs: 32 (broker) · 102 (isolation) · 78 (agent architecture) · 73 green-in-CI (workflows) · 157 (calendar) · 21 + 81 (orchestration).

Docs plus four docstrings. Verified docstring-only by AST comparison — worth noting the first two attempts at that check reported false positives, because subprocess.run(text=True) and open(encoding="utf-8") were decoding with different encodings, which made every em-dash compare unequal. The third run, with both sides decoded identically, is the one that holds.

Corrections the agents made to each other

Every one of the six caught something its own auditor got wrong: three extra stale anchors in WS-3, a fifth permission site in WS-8 that the auditor's replacement list had dropped, six wrong table names in WS-11 rather than four, _pending_user_input at executor.py:339 rather than :257 (the auditor's line was a comment mentioning the name), and — in WS-11 — that the "known Windows-only failure" is four tests, not one, three of them golden-eval trajectories. Anyone running that command with the audit's note in hand would have reported three regressions.

The board sweep then corrected two of those: calendar_ai_review.md is cited by migrations 92/97/100, not 98, and first_party does appear in six files — all comments, docstrings and a test helper, no field anywhere.

Deferred

• ai-company-brain/AGENTS.md's spec index has no calendar row at all and is missing agent_architecture.md; AGENTS.md:190 and apps/AGENTS.md:23 still carry the struck "broker ships with zero handlers" line. Filed as §5 residual 6.
• Two live docstrings in calendar.py cite migrations 93/94 for what shipped as 97/98 — recorded in the spec, not fixed (code change, out of a truth pass).
• calendar_ai_review.md and calendar_ux_review.md are unregistered sub-docs; ux_review is the sole home of the silently-dropped block-reminders deferral.

🤖 Generated with Claude Code