Lifecycle

This lifecycle is an optional OAT layer. Interop-only users can skip it.

OAT lifecycle order:

Discovery (oat-project-discover)
Spec (oat-project-spec)
Design (oat-project-design)
Plan (oat-project-plan)
Implement (oat-project-implement)
Review loop (oat-project-review-provide / oat-project-review-receive)
Summary (oat-project-summary) — generates summary.md as institutional memory; oat-project-pr-final and oat-project-complete auto-refresh it when missing or stale
PR (oat-project-pr-progress / oat-project-pr-final) — sets pr_open status
Revision loop (oat-project-revise) — optional; accepts post-PR feedback
Documentation sync (oat-project-document) — optional; reads project artifacts and code evidence to identify docs needing updates, checks for missing coverage of newly shipped capability areas, and auto-runs oat-pjm-update-repo-reference before scanning docs only when the PJM capability is available and this repository has adopted PJM
Complete (oat-project-complete)

Shortcut: oat-project-next reads project state and invokes the correct next skill automatically — use it instead of remembering which skill comes next. Complements oat-project-progress (which is read-only diagnostic).

Full spec-driven design supports three interaction modes: collaborative, selective collaborative, and draft-and-review. Selective collaborative drafts routine sections silently and presents high-risk or uncertain sections for live review. Quick-start lightweight design stays simpler and offers only collaborative or draft-and-review. See Design Modes for details.

Quick Look

What it does: explains the end-to-end lifecycle for tracked OAT projects, including alternate quick, lite, and import lanes.
When to use it: when you need the actual project execution model, not just a high-level overview of workflow mode.
Primary entry points: oat-project-new, oat-project-quick-start, oat-project-lite, oat-project-import-plan, oat-project-implement

Lifecycle Map

This map shows the spec-driven default. The quick, lite, import, and capture lanes each skip part of it — see Lane diagrams below. Lite in particular has no discovery, spec, or design step at all.

Post-implementation flow

Implementation closeout has one authoritative order:

Run final implementation verification.
Pass the mandatory final lifecycle review.
Resolve and disposition the configured oat-project-implement exit gate.
Run the configured pre-approval sequence.
Record final HiLL approval when the final phase is a checkpoint.
Run the configured post-approval sequence.
Mark implementation complete and emit the implementation success output.

The configured exit gate is independent from the mandatory lifecycle review, root-owned phase reviews, and optional oat_phase_review_gate. It must reach an allowed and fresh disposition before pre-approval work starts. A not_configured resolution is recorded explicitly as allowed/no_gate; it is not inferred from missing state, and a null, missing, or unrecognized resolver result fails closed as unresolved instead.

After implementation closeout finishes:

Summary (oat-project-summary) — generates summary.md as institutional memory from project artifacts; PR-final and completion will auto-refresh it if you have not already run it or if it is stale
Documentation (oat-project-document) — optional sync of project docs; decides whether repo-reference refresh should run before docs analysis using two independent checks, and should recommend new docs pages/directories when the shipped work introduces a capability area that the docs app does not already cover
PR (oat-project-pr-final) — creates PR description (auto-refreshes summary.md first when needed, then uses it as source), sets oat_phase_status: pr_open, and tracks actual PR existence with oat_pr_status / oat_pr_url
Revision loop (oat-project-revise) — accepts post-PR feedback:
- Inline feedback creates p-revN revision phases with prevN-tNN task IDs
- GitHub PR feedback delegates to oat-project-review-receive-remote
- Review artifacts delegate to oat-project-review-receive
- After revision tasks complete, state returns to pr_open
Complete (oat-project-complete) — accepts any phase status (pr_open, complete, in_progress), auto-refreshes summary.md before closeout when needed, and archives when selected by the workflow preference or completion prompt

Project-recap gate (non-lite)

The final-closeout orchestrator owns one project-recap gate. It runs after the final code review has passed and after any configured pre-approval summary and documentation steps, but before final HiLL approval. It neither replaces nor repeats the final review, and the stored order of the other pre-approval steps is preserved.

Recap intent resolves through oat-explainer-kit. A fresh project-recap manifest for the current completed implementation is reused rather than regenerated — fresh means it names recipe project-recap, belongs to this project, has a terminal outcome, and its recorded source hashes match the current approved inputs. A present-but-incomplete, wrong-recipe, or stale manifest does not qualify.

Recap outcomes are reported, not blocking: failed and built-not-durable are recorded as warnings and never block final approval, completion reporting, or later PR steps. The selected or attempted outcome and run path are included in the implementation completion report, and summary.md carries a single Explainer Outcome section when it exists.

Lite skips this gate entirely. Lite sets PROJECT_RECAP_REACHABLE=false and does not resolve recap intent, inspect recap runs, invoke oat-explainer-kit, or run the terminal-outcome guard. It proceeds from the required reviews through its stored optional steps to pr and sequence completion.

Lite mode collapses the default pre-approval sequence to PR creation only. oat-project-pr-final builds the body directly from the lite plan and implementation.md; summary and documentation remain explicit lite opt-ins, and retro is not added to the lite closeout sequence.

Completion archive behavior

On completion, OAT treats archive handling as an explicit closeout choice:

When archiving is selected, the local archive is written to .oat/projects/archived/<project>/.
Every completion that will actually archive — any durable (shared or synced) project with archive selected — keeps the activeProject pointer until the archive returns a validated receipt, and only then clears it. local projects are never durable and never archive, so they clear the pointer immediately, as does any completion that declines archive. An archive failure therefore leaves the pointer intact and the completion directly resumable. See Picking Up a Project for the post-archive resume and its manual recovery path.
When archiving is disabled or declined, durable projects remain at their active path. Synced completion still finalizes and pushes the project ref, commits the discovery record as complete, retains the checkout and ref, and attests a selected recap against the project-ref history.
For a synced project, closeout first finalizes the project artifacts and pushes them to refs/oat/projects/<project>. Archive then requires a clean, fully pushed checkout; copies it without the .git pointer or reviews/; seals configured durable exports; transitions terminal reachability to refs/oat/completed/<project>; removes the nested checkout; and deletes the tracked JSON record in the parent lifecycle commit. This order keeps SHA-pinned PR links valid without leaving active discovery state behind.
A completed-only ref and a completed ref plus same-SHA active alias are both valid terminal shapes. The active alias is inert: list omits it and pull/open reject it. Differing SHAs are a hard mismatch with recovery guidance.
If .oat/config.json enables archive.s3SyncOnComplete and sets archive.s3Uri, completion requires the S3 upload of a dated snapshot such as <archive.s3Uri>/<repo-slug>/projects/20260401-<project>/ to succeed before terminal record/ref cleanup. A failed configured upload leaves retry identity intact and does not claim closeout.
If .oat/config.json sets archive.awsProfile and/or archive.awsRegion, those values are forwarded to every aws invocation triggered by completion (preflight checks + aws s3 sync) and override any ambient shell AWS_PROFILE / AWS_DEFAULT_PROFILE / AWS_REGION / AWS_DEFAULT_REGION values. The repo's archive-scoped credentials are treated as deliberate intent so users don't have to unset shell env vars before running completion. See config-and-local-state for the full precedence chain.
If .oat/config.json sets archive.summaryExportPath, completion copies summary.md to <archive.summaryExportPath>/20260401-<project>.md.
Missing or unusable AWS CLI configuration blocks synced terminal cleanup when S3 durability is configured; when S3 sync is not configured, it is not part of the durability requirement.
oat repo archive sync can later sync all archived projects, or one named archived project, back down from S3; it selects the latest dated remote snapshot and materializes it into the local bare archive tree.
Persisted archive metadata binds the snapshot to the source-ref SHA, allowing an interrupted closeout or legacy complete record to resume without rematerializing an active project. Dashboard and list surfaces omit fully retired projects and report precise cleanup or ref-mismatch diagnoses for remaining legacy state.
oat project links continues to render full-SHA links through the completed ref. oat project prune is a separate destructive choice that removes terminal ref reachability but preserves local and S3 archive snapshots.

Absorbed-project retirement sweep

Retiring a project that a later project absorbed is a semantic claim, not a filesystem operation. Deleting or archiving the old scaffold does not make the roadmap stop promising its work, so completion checks the claim against the prose that still makes it.

When a quick project consolidates earlier scaffolds, quick-start records absorbed_projects and absorbed_backlog_ids in that project's state.md frontmatter. Those two fields are the only inputs the sweep takes.

Recording them is a quick-mode step today. A lite project that consolidates earlier scaffolds records nothing, so the sweep finds no inputs and degrades to its recorded note rather than checking that project's surfaces.

Inside the project-log completion gate — after the log status probe and before the roll-up and seal — completion searches the active planning surfaces for each absorbed slug and backlog ID, and for future-oriented ownership language tied to them:

.oat/repo/pjm/roadmap.md (the Now/Next/Later lanes and the sequencing map)
.oat/repo/pjm/current-state.md
the Curated Overview in .oat/repo/pjm/backlog/index.md
the scope-nested state.md of projects that are still active, under the configured projects root, skipping the archived tree and any project already recorded as complete

The sweep is advisory. A raw match is never a hard block on closeout: each hit becomes a named finding carrying a recorded disposition: fixed in the same run, accepted as historical with a reason, or — in an autonomous run — recorded as deferred advisory. Prose that clearly describes past state is exempt, so a dated history entry, a retro, a decision record, or a changelog line reads as evidence rather than a stale claim.

Placement is the point. Dispositions are appended to the project log before the roll-up, so the roll-up summarizes them and the completion seal remains the final entry — no project-log append ever follows the seal. Three variations change where and how the dispositions are recorded: an autonomous run records each finding as a deferred advisory warning and continues, a project with no log reports them in the final completion summary, and a resumed completion whose log already carries a seal runs report-only and appends nothing.

A resume detects that seal from oat project log check, which reports a sealed field, rather than by re-reading the log. The invariant is enforced in the CLI: replaying the seal reports already-appended and leaves exactly one seal entry; a replay recognized by its own idempotency key reports already-appended even on a sealed log (nothing is appended), and every append carrying new content is refused with status: "sealed" and a non-zero exit.

Repositories that never adopted PJM skip the sweep with a one-line note. A missing planning surface degrades the sweep, never the closeout.

Phase status: `pr_open`

After oat-project-pr-final runs, state.md shows oat_phase_status: pr_open. This signals:

The project is in its post-PR review posture
The project is NOT done — agents should not start a new project
Next steps: oat-project-revise (for feedback) or oat-project-complete (when approved)

Actual PR existence is tracked separately from oat_phase_status:

oat_pr_status records whether the PR is merely ready to create, open, closed, or merged
oat_pr_url records the tracked PR URL when a PR exists

This distinction matters during completion: oat-project-complete can skip the "Open a PR?" prompt when oat_pr_status: open is already present.

Auto-review at HiLL checkpoints

Approval-aware post-implementation sequencing

workflow.postImplementSequence can use the legacy string values or a structured { preApproval, postApproval } value. After final verification, final review, and an allowed, fresh implementation exit-gate disposition, OAT snapshots the effective sequence, runs ordered pre-approval steps, records final HiLL approval, and only then runs post-approval steps. The snapshot is restart-safe: an incomplete sequence routes back to implementation and resumes from its first incomplete step.

Structured sequences accept summary, document, pr, and retro. retro is post-approval only: placing it in preApproval invalidates the structured value. This placement lets the retrospective include the final approval and feedback tail while still running before project completion freezes lifecycle artifacts. The legacy string mappings are unchanged and do not add a retro step.

When a pending post-approval retro step runs, OAT dispatches oat-project-retro in generate mode. Applying repo improvements and filing tracker items remain separately consented through interactive confirmation or workflow.retro.* configuration.

oat-project-next checks oat_implement_exit_gate before every normal post-implementation route. Missing, pending, blocked, malformed, or stale state routes back to oat-project-implement even when oat_phase_status is complete or pr_open. Only an allowed, fresh disposition can continue to summary, documentation, PR, or project completion.

Retrospective completion safety net

Before an interactive completion closes the project, oat-project-complete checks for {PROJECT_PATH}/references/project-retro.md. If the artifact is missing, it offers to generate one before completion. If the artifact exists, completion does not offer another retro; it may note unsettled promotion or filing registers. Non-interactive completion skips this offer, so autonomous generation occurs only through an explicitly configured post-approval retro step.

When workflow.autoReviewAtHillCheckpoints is enabled or plan.md frontmatter sets oat_auto_review_at_hill_checkpoints, completing a HiLL checkpoint automatically runs the extra lifecycle review scoped to every implementation phase not already covered by a passed whole-phase code review, through the just-completed checkpoint. Mid-implementation multi-phase reviews use inclusive phase-range scopes such as p02-p03; the final implementation checkpoint uses code final. The review uses auto-disposition mode (minors auto-converted to fix tasks, no user prompts). Disabled by default. Legacy autoReviewAtCheckpoints and oat_auto_review_at_checkpoints are still read as fallbacks. This does not control Tier 1 per-phase oat-reviewer gates.

Phase-review setup during planning

Spec-driven, quick, and import planning run one shared setup after stable phase IDs exist and before the plan artifact review. Lite omits this planning-time setup and prompt, while its built-in implementation phase and final reviews remain required. The target probe qualifies only an explicitly configured, enabled, and available review target, then offers all phases, selected phases, or disabled. Existing explicit oat_phase_review_gate values are preserved unchanged without re-prompting. Probe failure, no qualifying target, non-interactive execution, or user decline leaves phase review disabled. Provider native plan mode inherits this behavior through the import-plan lane.

Implementation modes

oat-project-implement v2.0 dispatches one subagent per phase (not per task). Capability detection at skill start selects a tier, locked for the run:

Tier 1 (subagents): native subagent dispatch via Claude Code, Cursor, or Codex with spawn authorization.
Tier 2 (guarded sequential execution): orchestrator reads the agent files and uses a target-preserving route when native subagents are unavailable or authorization is declined.

A concrete managed reviewer remains bound across both tiers. Codex uses the exact registered role or a child pinned to the resolved model and effort; Claude passes the exact resolver-returned dispatchArgs.model; Cursor launches the exact providers.cursor.dispatchArgs.variant native reviewer variant. Retries preserve that complete target. Tier 2 does not authorize a target downgrade. Inline review is allowed only with verified equivalent host controls, or for explicit inherit/default behavior or the documented managed-uncapped reviewer base-role exception; otherwise the review blocks.

Within either tier, parallelism is expressed as plan metadata:

Sequential (default): plans with no oat_plan_parallel_groups field, or with an empty array. Phases run in plan order on the orchestration branch.
Parallel groups: phases listed together in oat_plan_parallel_groups run concurrently in worktrees (Tier 1 only) and merge back to the orchestration branch in plan order. Groups themselves execute sequentially.

See Implementation Execution for the full execution model — tier detection, bounded fix loop, fan-in, merge-conflict handling, dry-run, and resumption.

Review receive behavior

oat-project-review-receive now presents a findings overview before asking for any disposition decisions.
Findings are shown with stable IDs by severity (C*, I*, M*, m*) so follow-up choices map clearly to specific items.
For each finding, the receive step summarizes the reviewer note, adds agent analysis, and gives a recommendation (convert now vs defer with rationale).

Alternate lifecycle lanes

Quick lane diagram

oat-project-quick-start (adaptive discovery — provide a project name and optional description; if only the name is provided, quick-start asks for the missing description before discovery. Well-understood requests synthesize quickly, exploratory requests invest in solution space exploration. Before scaffolding, the skill checks inherited git state and asks whether to commit, proceed, or abort when the worktree is already dirty.)
Decision point: straight to plan, optional lightweight design.md, or promote to spec-driven
Implement: oat-project-implement (sequential by default; parallel when oat_plan_parallel_groups is declared)
oat-project-review-provide / oat-project-pr-final
Optional oat-project-promote-spec-driven to backfill spec-driven lifecycle artifacts in-place

Lite lane diagram

oat-project-lite runs one batched interview, authors a single-phase plan.md with validation criteria, and pauses once for approval
Implement: oat-project-implement runs the single phase without HiLL checkpoint prompts
Pass the mandatory final review, then route directly to oat-project-pr-final
Optional oat project promote <project-path> --to quick when the work no longer fits one sitting

Import lane diagram

oat-project-import-plan (checks inherited git state before import scaffolding so sync-generated or unrelated dirty files do not silently roll into project bookkeeping)
Implement: oat-project-implement (sequential by default; parallel when oat_plan_parallel_groups is declared)
oat-project-review-provide / oat-project-pr-final
Optional oat-project-promote-spec-driven to switch project mode to spec-driven lifecycle

Lane diagrams

Spec-Driven workflow lane

During the Design step, oat-project-design asks how to work through the document unless a mode was selected by argument, environment, or workflow.designMode. The three full-design choices are collaborative, selective collaborative, and draft-and-review.

Before a plan is marked ready for implementation, planning resolves the current provider's dispatch policy from workflow.dispatchPolicy.mode and workflow.dispatchPolicy.policy, or from project state.md frontmatter. The legacy workflow.dispatchCeiling.providers.<provider> config and oat_dispatch_ceiling frontmatter remain compatibility inputs for capped managed policies. If no policy is configured and the session is interactive, planning asks once and stores the answer as oat_dispatch_policy; non-interactive planning with an unresolved policy is not implementation-ready and blocks at the planning boundary until the resolver succeeds. Implementation preflight retains the same defensive check, but it is not the first place an unresolved plan fails. Uncapped is explicit managed selection with no maximum cap; Inherit Host Defaults is separate and means OAT does not select model or effort controls.

Quick lane

Quick lane lightweight design intentionally keeps a smaller collaborative/draft choice. Selective collaborative becomes available only after promotion into the full spec-driven design lane.

Quick-start follows the same dispatch policy rule at the plan boundary: capture the policy when interactive, persist oat_dispatch_policy to project state, and avoid any mid-implementation policy prompt. Legacy dispatch-ceiling state is still read as compatibility capped-policy input, but new quick-start selections should use the dispatch-policy names, including explicit Uncapped and Inherit Host Defaults.

Lite lane

Lite keeps planning and validation in plan.md. The interview selects a minimal, product, technical, or both content shape and records the rationale. Product Behavior is required for user-visible changes; Technical Design is required for cross-module, data/state-format, or consumed-contract changes. It has no separate discovery, spec, or design artifact, skips implementation HiLL prompts, and routes a passed final review directly to PR creation. Summary and documentation run only when enabled through the lite-specific post-implementation sequence.

Each task chooses evidence proportionate to its risk instead of requiring one test-first recipe. Behavioral changes need a proof that fails without the change; refactors default to characterization-first when existing coverage is insufficient; documentation-only work may use formatting, link, spelling, or build checks. UI changes require visual proof. In autonomous runs, available computer-use capability performs that proof; otherwise execution stops at an explicit proof boundary.

Promoting a lite project to quick

When lite work outgrows one sitting, oat project promote <project-path> --to quick converts it in place. Each artifact has a different fate: the authored lite plan is moved to references/lite-plan.md, a new discovery.md is derived from its five core sections plus any adaptive Product Behavior and Technical Design, a fresh quick template takes over the plan.md path, and state.md is rewritten to quick mode.

The project slug, directory, and branch are untouched, so history and any open work continue uninterrupted. oat_ready_for is stamped on both discovery.md and state.md; the quick-start recommender reads artifact readiness, so consumers must preserve both. Promotion refuses rather than half-applying: a non-lite project, an existing references/lite-plan.md, invalid authored sections, an unresolved scope, or an unreadable template each return a stable categorical reason and exit non-zero. See CLI Reference for the full refusal contract.

Import lane

Capture lane

Retroactive project creation for work done outside the OAT project workflow. Common scenario: mobile/cloud sessions where you brainstorm and implement with an agent, then want to open a PR and review from your desktop.

Entry point: /oat-project-capture (skill-only, no CLI command — requires agent conversation context).

Key differences from other lanes:

No plan generation — the work is already done; the scaffold-created plan.md template is kept but not authored
Discovery from conversation — discovery.md captures intent and decisions from the agent's conversation context, not from requirements analysis
Implementation from commits — implementation.md is populated from commit history with SHAs, not from executing plan tasks
Lifecycle state is user-chosen — user decides whether the project is ready for review or still in progress

Artifact progression

discovery.md -> spec.md -> design.md -> plan.md -> implementation.md -> summary.md

Quick lane progression:

discovery.md -> [design.md (optional lightweight)] -> plan.md -> implementation.md

Lite lane progression:

plan.md -> implementation.md -> pr/*.md (summary.md and documentation are opt-in)

Import lane progression:

references/imported-plan.md -> plan.md -> implementation.md (spec.md/design.md optional)

Capture lane progression:

discovery.md (from conversation) + implementation.md (from commits) — no forward-looking artifacts

Operational rules

Keep state.md, plan.md, and implementation.md synchronized.
For a synced project, publish lifecycle artifact writes with oat project push; do not stage .oat/projects/synced/<project>/ on the parent branch.
On arrival, run oat project pull before reading a synced project's state.
Stop at configured HiLL checkpoints.
Do not move lifecycle forward when required review gates are unresolved.
Project entry skills (oat-project-new, oat-project-quick-start, and oat-project-import-plan) surface inherited dirty git state before scaffolding. If the dirty list includes .oat/sync/manifest.json, .claude/, .cursor/, or .codex/, the skill calls out that those paths are typically sync output and offers Commit now, Proceed anyway, or Abort.

Reducing lifecycle friction with workflow preferences

The lifecycle has several interactive prompts that power users often answer the same way every time — HiLL checkpoint behavior, archive on complete, auto-create PR, post-implementation chaining, final review execution model, and re-review scope narrowing. These can be configured once via workflow.* preference keys and respected automatically by skills.

See the Workflow preferences section in the Configuration guide for the full list of keys and how to set them. Preferences resolve through a three-layer chain (env > repo-local > repo-shared > user > default), so you can set personal defaults at user scope once and override per-repo only when needed.

Active project resolution

Active project state is stored in .oat/config.local.json (activeProject, repo-relative path).
Projects root is stored in .oat/config.json (projects.root) and can be read via oat config get projects.root.
Workflow skills prefer oat config get activeProject / oat config get projects.root rather than reading pointer files directly.

Capability availability versus repository adoption

Lifecycle skills that touch repository project-management state make two separate checks, and both must hold before any PJM write:

Capability availability — oat tools has project-management answers whether the PJM skills and templates are installed and complete at project or user scope. Since packs default to user scope, this is usually satisfied by a user-scope install with no repository footprint.
Repository adoption — oat pjm doctor --json answers whether this repository adopted PJM. Branch on its adoption.state field: declared and inferred-legacy allow repository PJM writes; partial-initialization and none stop with oat pjm init as the recovery.

Pack presence is never treated as evidence of repository adoption. A skill that finds the capability available but the repository unadopted reports the actionable oat pjm init stop instead of scaffolding implicitly. See Install vs. initialize.

Brainstorming integration with the project lifecycle

oat-brainstorm (in the brainstorm tool pack) interacts with the project lifecycle in two distinct ways: it can seed a brand new project from a brainstorming conversation when no project is active, or fold back into an active project when one is.

Seeding a new project from a brainstorm

When the brainstorming destination is "promote to a new OAT project", the dispatcher confirms a project slug and one of three modes: lite, quick, or spec-driven. For Quick and Spec-Driven projects, it runs oat project new <slug> --mode <mode> and writes discovery.md directly from the brainstorming payload — Initial Request, Solution Space with approaches considered, Chosen Direction, Key Decisions, and Open Questions. It writes only discovery.md (never a partial design.md), then points to oat-project-quick-start or oat-project-design without auto-chaining. For Lite projects, it instead seeds the five sections in plan.md, leaves discovery.md and design.md absent, and points to oat-project-lite.

This is a parallel entry path into all three lanes, not a new lane. Quick and Spec-Driven users arrive at oat-project-quick-start or oat-project-design with discovery already populated; Lite users arrive at oat-project-lite with the single-phase plan already seeded.

Folding back into an active project

When oat-brainstorm fires while a project is active, it offers a 3-way picker before any pack-filtered destination shows up:

Related to the project → for Lite, fold the synthesis into plan.md immediately above ## Phase 1 and hand off to oat-project-lite; for Quick and Spec-Driven, use the most-specific upstream artifact (design.md if it exists, otherwise discovery.md) and hand off to oat-project-quick-start or oat-project-plan. Any mode with an open PR hands off to oat-project-revise. The fold-back commit is safety-gated: a preflight git status --porcelain -- "$ARTIFACT_PATH" checks for dirty state, staging is exactly git add -- "$ARTIFACT_PATH" (never -A, never globs), and the handoff prompt only prints after the scoped commit succeeds.
Independent of the project → the active project is acknowledged but doesn't constrain the picker; brainstorm routes through its standard pack-filtered terminal-state options (new project, backlog item, idea, doc-to-path, inline).
Related but supplementary → write a brainstorming reference file at .oat/projects/<scope>/<project>/brainstorming/YYYY-MM-DD-<topic>.md, alongside pr/ and reviews/. No lifecycle artifact is touched.

The fold-back path is what makes "we got to plan and realized the design missed something" recoverable mid-project: the upstream artifact gets updated and committed cleanly, then the user re-runs plan authoring with a "don't refresh, integrate" context that preserves review-table state.

Reference artifacts

.oat/projects/<scope>/<project>/spec.md
.oat/projects/<scope>/<project>/design.md
.oat/projects/<scope>/<project>/plan.md
.oat/projects/<scope>/<project>/implementation.md

On this page