---

type: adr id: adr-0060 status: accepted created: 2026-06-11 updated: 2026-06-11

ADR-0060 — The corpus Workspace: hybrid layout, committed flow artifacts, the review packet

Context

The product's wedge is review-by-exception: turning large agent output into requirement coverage, evidence, and a human-attention list. Agent-authored summaries are not trustworthy on their own — evaluators favor their own generations [SELFPREFER], judges carry position/verbosity biases [JUDGEBIAS], and unsupported "done" claims are a recurring failure (small-N, preliminary) [EVIBOUND]. The external survey found no tool ships a persisted, independent, exception-routing review packet (the verified gap — never claim "nobody reviews agent output"); it also found every competing tool keeps specs in-repo, while external requirements stores exist at the RFC/requirements-repo granularity enterprises already run, with drift/discoverability as the known failure mode of separation (recorded counter-sources: the spec-kit issue threads #1191/#876/#1059 and the Fiberplane "Drift" post — survey V-059/V-024).

Decision

1. Workspace layout (hybrid). Durable intent lives in feature folders: specs/<feature>/spec.md plus co-located supporting docs (audit, research, prd, rfc, inventory). Flow artifacts live in committed type folders: intake/, tasks/, reviews/, findings/, inventory/, change-plans/ (the latter two appear when the work needs them). Project decisions in decisions/ (numbered ADRs). A hand-edited status.md workboard at the root. .agents/ holds only tooling. Both naming depths are valid (folder-per-artifact with NNN- prefixes, or flat files for small projects).
2. Flow artifacts are committed. Tasks, reviews, and findings are durable workspace content — the review packet that links its PR is the default record of work. (A PR-only record remains the floor for teams with no workspace.)
3. Co-located and external are co-equal defaults. A single-repo team keeps the same tree inside its repo (optionally under a visible corpus/ directory). The external workspace is framed as a Git-native, agent-readable form of the external requirements store larger organizations already run — never as the common practice of comparable tools (it is not; the in-repo norm and the drift failure mode are the recorded counter-evidence, survey V-021/V-024). One spec store can govern many code repos.
4. The review packet (template frozen here, shipped at starter-kit/templates/review.md): frontmatter type: review, id, task, pr, status: draft|pass|blocked|needs-human; sections Summary / Changed files / Requirement coverage table (ID | Result | Evidence | Human attention, results Pass · Fail · Unverified · Blocked) / optional Change-plan coverage table (same columns) / Human attention / Suggested decision. Hard rules: a Pass needs pasted output or a CI link; an empty Evidence cell means Unverified, never Pass (checklist level); reviewers spot-check at least one green row's evidence (automation-bias countermeasure — convention level); the exception-trigger list the packet routes is: unverified/failed requirements · out-of-scope changes · risky files · missing test output · changed public interfaces · DB migrations · security-sensitive changes · new finding candidates · blocked questions.
5. The task packet (template frozen here, shipped at starter-kit/templates/task.md): frontmatter type: task, id, source[] (spec and/or change plan), scope[] (requirement ids), status; sections Source / Scope ("Implement or preserve") / Do not change / Affected areas / Verify checklist / Agent instructions (read sources first; stay in scope or stop and say why; run every Verify item and paste real output; re-read your own diff as a skeptic before finishing; leave a summary of changed files, commands run, findings) / Findings.
6. Spec evolution is routine. A spec is amended in place after review feedback (edit the requirement, keep its ID, note material drops under "Dropped from sources"); workspace↔code drift is detected at review time by the packet's coverage table, and stale specs are flagged on the status board. No regeneration ceremony.

Alternatives considered

Alternative	Why weaker
Gitignored task/review scratch	Discards the wedge's record; the packet is the product's durable evidence
Pure per-feature folders for everything	Flow artifacts cross features (one review covers one task, not one feature); type folders keep the board scannable
External-store as the sole default	Contradicts the documented in-repo norm and adds drift risk for small teams

Consequences

Positive: requirement → evidence is one folder hop; the board reads at a glance. Negative: committed flow artifacts grow the workspace (archive convention left to teams). Neutral: PR remains the merge mechanism.

Status

Accepted. Supersedes ADR-0004; partially supersedes ADR-0052 (feature folders kept for intent; the scratch rule and memory-dir home are replaced); refines ADR-0049, ADR-0050, ADR-0032, ADR-0030.

Propagation

Templates (task, review, status), docs/03/05/06/08, ADOPTING, kit shell, conformance task/review schemas, examples, evals.

Addendum (2026-06-11): the worker's run record folds into the review packet (the run summary

• evidence cells); a standalone trace artifact exists only as a reserved machine-record sketch on the future-CLI page. Inventories live in the type folder inventory/; the feature folder co-locates only spec-supporting documents (audit, research, prd, rfc).

Ledger note (2026-06-12): the workspace layout this ADR prescribes is shipped pre-built by the starter kit per ADR-0069.

Ledger note (2026-06-12, later): the addendum's run-record clause ("the worker's run record folds into the review packet") is superseded by ADR-0072 — the run summary lives in the task packet as a digest; the task and review packet formats are amended additively by the same ADR.