Poppy TeamPT
Poppy Team

English

Português

English

Português

Appearance

PT
PrumoProtocol and CLIAbout the projectRepository ↗

Static copy

Copied from docs/en/governance/m6-exit-gate.md in https://github.com/poppy-team/prumo (MIT). Pinned to revision e91694d3959be1ed92757b34063efe0b2191821f, blob 0ae8151ecf7db725da0e5bdeac147fb1e2a33fbb. The source repository remains canonical; this copy is refreshed through a sync pull request, not live.

M6 Living Plan / Interview Engine — Exit Gate ​

Status: COMPLETE (M6 Exit Gate Passed)

Exit gate criteria ​

M6 exit gate (per docs/development/phases.md): a new project can go from initial intent to implementation-ready via interview without a manual megaprompt; Goal-specific planning closes only relevant gaps; decisions/open questions carry authority and provenance; resume does not depend on transcript; docs/readiness/governance feedback loop works; question/decision evals reach the approved baseline; no agent suggestion is silently promoted.

1. Initial intent → implementation-ready without a manual megaprompt ​

  • Deterministic end-to-end dogfood test TestDogfoodZeroToReady (cmd/prumo/plan_z2r_test.go) drives the full loop through run(): blocked readiness → questions → answered decisions → proposed delta → governance patch → applied delta → passing readiness → blueprint proposal.
  • Committed sample project examples/living-plan-sample/ ships at the ready state with a replayable seven-step recipe (README.md).
  • Verified on the sample with the real CLI: docs readiness --goal G-SAMPLE transitions from {"ready": false, "blocking_contracts": ["project.scope"]} to {"ready": true, "coverage": [[product.vision, implementation-ready], [project.scope, implementation-ready]]}.

2. Goal-specific planning closes only the relevant gaps ​

  • Planning and readiness are goal-scoped: prumo-agent plan --goal <id>, docs readiness --goal <id>; the documentation engine ignores unrelated contracts when computing coverage for a Goal (M5 regression suite).
  • The sample loop only touches product.vision and project.scope; unrelated contracts stay untouched.

3. Decisions / open questions carry authority and provenance ​

  • Decision proposals expose class, actor/source, authority, confidence (inference-only), status, and affected contracts/docs (internal/planning/decision.go, resolver internal/planning/resolver.go, E-G02/E-G03).
  • Open questions registry tracks blocker status, linked contract, owner, created/resolved/superseded state, and resolution evidence (internal/planning/question.go, E-G01).
  • A resolved question is not deleted; it remains traceable to its decision.

4. Resume does not depend on transcript ​

  • Session checkpoint persisted under .ai/plan/sessions/<id>.json as session.Checkpoint() (internal/app/sessionstore.go, E-G07).
  • prumo-agent plan resume recompiles context from canonical decisions + open questions + checkpoint (internal/app/resume_test.go, internal/planning/session_test.go); no transcript is reintroduced.

5. Docs/readiness/governance feedback loop works ​

  • Loop proven by TestDogfoodZeroToReady and the sample README steps: readiness → questions → answered decisions → delta propose → repository governance authoring → delta apply (records decision evidence) → readiness recompute → goal/plan proposal.
  • plan delta [--apply] never writes canonical docs; governance authoring + README documents that flow (E-G05/E-G09).

6. Question/decision evals reach the approved baseline ​

  • The approved design (Notion §70.18) defines the baseline as deterministic conversation evals; live-model evals belong to the future Harness Eval Suite and are explicitly not a unit-test requirement.
  • Every listed deterministic scenario has a regression test: blocker sorted before nice-to-have; explicit decision over inference; agent suggestion not auto-accepted; locked-decision conflict produces a finding (E-G04); resolved question updates the registry; unrelated docs gaps ignored for Goal readiness; resume preserves accepted decisions without transcript.
  • Run in CI via go test -race ./....

7. No agent suggestion silently promoted ​

  • agent-suggestion classifications never promote to decisions and keep the question open (answer authority enforcement in internal/planning + cmd/prumo/plan_z2r_test.go: TestPlanAnswerNeverPromotesAgentSuggestion, TestPlanAnswerUnresolvedKeepsQuestionOpen).

Evidence (framework dogfood on this repository) ​

  • examples/living-plan-sample/ — committed sample, ready state, replay README, and the applied delta record docs/governance-delta-applied.json.
  • Full CLI surface exercised against the sample: plan questions, plan answer, plan decisions, plan delta [--apply], plan blueprint [--plan], docs readiness --goal (E-G08 JSON envelope + text modes).

Test evidence ​

  • go test -race ./... — pass across all packages.
  • go vet ./..., gofmt -l . — clean.
  • pytest -q — 127 pass (Python v0.3 oracle preserved).
  • M5/M6 regression suites: coverage/re readiness isolation, delta lifecycle, decision preview contradiction findings, authority resolution, checkpoint round-trip, resume context compilation, command-line JSON envelopes.

Governance ​

  • M6 merged to main via squash PRs: E-G01 (#24), E-G02 (#25), E-G03 (#26), E-G04 (#27), E-G05 (#28), E-G06 (#29), E-G07 (#30), E-G08 (#31), E-G09 (#32).
  • Canonical spec: docs/runtime/living-plan.md; samples/dogfood under examples/living-plan-sample/.

Excluded from M6 exit gate ​

  • Live-model conversation/decision evals remain a Harness Eval Suite item (Notion pages 61 and 70 §18); deterministic regression evals already satisfy the gate.
  • Adoption Engine (M7), traceability (M8), and Experience (M9) are explicit non-goals for M6.