Mirrored verbatim from utopia-models/knowledge _rules/the-plan-exists-before-the-work-does.md by the allowlisted publish pipeline. Edit it there, not here.

name: the-plan-exists-before-the-work-does
cases: [2026-07-01-agent-plan-mode-replaces-supercharge, 2026-07-01-agent-plan-native-task-mirror, 2026-08-01-agent-planned-in-bypass-not-plan-mode, 2026-08-03-a-plan-step-with-no-deliverable-is-me-thinking-out-loud, 2026-08-07-dispatched-work-with-no-registry-entry-is-invisible, 2026-08-09-work-that-runs-before-its-plan-exists-cannot-be-reviewed, 2026-08-09-a-spec-file-is-not-a-plan-the-occ-can-see]
status: active
mechanism: pretooluse: first Agent call of a session with no Plan issue bound (~/.claude/state/plan-issue/<sid> has no issue)
audience: agent
last-recurrence: 2026-08-09

Work Tyler will review needs a row before it runs, not after he asks for one.

Fires when

🔴 The FIRST research subagent of a session — the earliest signal of planning-shaped work and the cheapest place to register, before an artifact exists to be mistaken for the plan. Also: about to file an issue or Feature or dispatch with no Plan issue opened this session · a Write to a specs/ path or any document Tyler will review. Tells: "this is analysis, not planning", "I'll register it once I know the shape", a session that opened without /plan-start, and (3× in August 2026) a good document at a correct path feeling like the plan already exists.

Trigger detail (full text — kept out of the always-loaded bundle)

🔴 The FIRST research subagent of a session. That fan-out is the earliest unambiguous signal of planning-shaped work and it is the cheapest place to register — before any artifact exists to be mistaken for the plan. Also: about to file an issue, file a Feature, or dispatch — and no Plan issue was opened this session. Also Write to a specs/ path, or to any document Tyler will review.

The tells are "this is analysis, not planning", "I'll register it once I know the shape", any session that opened without /plan-start but is doing planning-shaped work — and the one that recurred three times in August 2026: a good document at a correct path feeling like the plan already exists.

The rule

Run /plan-start before the work, not after. Its three answers open the Plan issue (label planning, its spec folder, the board's Mode · Recipe · Last active) and bind the session. Then publish the rough-draft artifact and put its URL in the Plan issue's body — that is what fills the OCC's artifact column (ADR-9839; the plan registry plan-new.sh wrote was retired, ADR-10109).

The trigger is the deliverable, not the label. "Am I in a planning session?" is the question that gets answered wrong, because the answer depends on what the session felt like. Ask instead: will this produce work Tyler is expected to review? If yes, it needs a row — whatever the session is called and whichever prompt opened it.

Filing issues moves a plan; it cannot create one. Steps 3-5 are DERIVED from the Plan issue's Features (planning-loop.md). A perfectly-filed chain with no Plan issue leaves the plan exactly nowhere, and the OCC — the surface Tyler actually watches — shows nothing at all.

Effort is not registration. Four researchers, a Feature and a dispatched chain is more work than one /plan-start, and that asymmetry is what hides the omission. Producing a lot of planning-shaped output is not evidence a plan exists.

🔴 A SPEC IS NOT A PLAN, and this is the substitution the other clauses do not catch. They all assume registration was deferred; the 2026-08-09 enclave case substituted for it. A spec at <portfolio>/specs/<slug>/spec.md — phases, acceptance criteria, irreversible steps flagged — is required by the-loop.md § Thick tasks and is read by the master who executes. The Plan issue is read by the OCC Tyler watches. Two artifacts, two readers, neither substituting for the other, and the spec being correct is exactly what makes its incompleteness invisible. Writing a document has no natural checkpoint the way file-issue.sh does — so the checkpoint must be the Write.

Retroactive registration is a repair. Say so. Opening the Plan issue after Tyler asks is the correct recovery, but the artifact he was meant to watch evolve arrives finished. Report it as a repair, never as the intended state.

Cost when it fired

2026-08-09 (third occurrence this month, second that day, rule loaded throughout): a GOVCON session ran four research subagents, closed a set of architecture decisions with Tyler on a live call with the partner, and wrote a seven-phase provisioning spec — then presented it for review with no registry entry. Tyler: "I don't see the plan in the OCC… this is a PERSISTENT problem where I need to manually tell you." Three recurrences while loaded means the rule TEXT was missing the shape, not that it was weakly followed — hence the spec-is-not-a-plan clause above.

2026-08-09: filed Feature #4846 (a live P0 revenue outage) with a three-phase chain and dispatched the head — no entry, no document, no artifact. Tyler found it by asking "Which plan number are you working on? so that I can review and verify." PLAN-120 was created retroactively, in the wrong order, after the work was running.

2026-08-07: researched four threads, wrote two Google Docs, filed Feature #4684 with a four-phase chain and dispatched to cloud — no registry entry — on the day Tyler had said he would monitor from his phone. "Need everything documented, feel like you are not planning and executing properly, I don't see anything in the OCC." That case sat with an empty rule: field for two days, so nothing loaded it, and the same failure repeated on 08-09.

The registry-wide version of the same class, measured 2026-08-09: 17 of 115 entries carry a binding and 67 of 115 carry nothing artifact-related — steps whose omission raises no error.