corpus Reference

Works today — plain markdown plus your agent; no corpus tooling required.

One page of lookup tables: labels, results, verification methods, artifact homes, the loop, and the check ids. Definitions live in the linked pages; the exact file formats live in the kit templates (templates/) and are never restated here.

The loop

Pull → (Inventory) → Spec → (Change Plan) → Task → Run → Review → Close
        brownfield          structural

Step	What it produces
Pull	an intake file — the upstream ticket, captured verbatim
Inventory (conditional)	a map of existing code before structural/brownfield work
Spec	intent, non-goals, requirements with Verify with: lines
Change Plan (conditional)	how the codebase changes while named behavior provably survives
Task	a bounded task packet an agent can run
Run	the change, on its own branch, plus an agent run summary
Review	a review packet — requirement coverage, evidence, exceptions
Close	merge or block, board update, findings saved

Not every task needs every step — the skip rules and per-shape flows are in the basic workflow.

IDs and labels

Prefix	Artifact	Example
SPEC-	spec	SPEC-auth-refresh
TASK-	task packet	TASK-auth-refresh-1
REVIEW-	review packet	REVIEW-auth-refresh-1
FINDING-	finding	FINDING-rotated-tokens
INV-	inventory	INV-billing
CHANGE-	change plan	CHANGE-billing-split

Requirement labels inside a spec:

Label	Meaning	Where
AC-NNN	acceptance criterion — a verifiable requirement	any spec
C-NNN	constraint — restricts how requirements may be met	SOL form
I-NNN	invariant — a property that must keep holding	SOL form
IF-NNN	interface — a declared boundary	SOL form
Q-NNN	open question, optionally [blocking]	SOL form

A file's identity is its frontmatter type:, never its filename — see artifact formats. SOL is the optional structured-requirements notation (format: sol) — see structured requirements.

Review results

One result per requirement row in the review packet (internally: verdicts).

Result	Meaning
Pass	Verified — pasted output, a CI link, or a named human's recorded observation (manual method) sits in the Evidence cell.
Fail	Verified, and the requirement is not met.
Unverified	No evidence. Every empty Evidence cell reads as Unverified, never Pass.
Blocked	Cannot be verified until something else is resolved.

Lifecycle results, used in advanced workflows to qualify an earlier result:

Result	Meaning
Waived	A Fail or Unverified explicitly accepted by a named human, with reason and expiry.
Stale	A prior Pass whose evidence predates a change to the requirement or the code it exercised — see drift.
Contradicted	Two pieces of evidence disagree, or the code disagrees with the requirement.

The evidence rules and exception triggers are in reviewing output.

Verification methods

What kind of evidence backs a requirement's Verify with: line (internally: proof types).

Method	One line
test	An executable test exercises the behavior (unit/integration/e2e are scopes of this one method).
static	Type-check, lint, build, or static analysis over the code as written.
contract	A boundary check against a declared interface — API schema, consumer contract.
property	A property-based check over generated inputs; strong for invariants.
model	An exhaustive or model-checked argument; the strongest, and the rarest.
perf	A measured benchmark against a stated budget.
security	A scanner or a targeted security test against an attack class.
manual	A named human checked it and recorded what they saw — always with a reason and the observation itself.
monitor	A production probe or alert observes the behavior after merge.

Artifacts and where they live

Artifact	Tier	Home in the workspace
intake	core (recommended for tracker-originated work)	intake/
spec	core	specs/<feature>/spec.md
task packet	core	tasks/
review packet	core	reviews/
finding	core	findings/
status board	core	status.md
inventory	core when the work is brownfield	inventory/
change plan	core when the work is structural	change-plans/
adr	advanced	decisions/, numbered
audit · bug · research · rfc · prd · threat-model · release-note	advanced	co-located in specs/<feature>/

The full layout, both naming depths, and the code-repo boundary are in where files live; every format is catalogued in artifact formats.

Checks — the quick list

Common mistakes to check for in a spec — plus the two change-plan checks. Level: checklist today; toolable — corpus-cli's corpus check implements this list. Full descriptions, the honesty legend, and the SOL catalogue are in checks.

ID	Name	Severity
C001	unique-ids — every requirement ID appears exactly once	hard error
C002	duplicate-id — no other file claims the same id:	hard error
C003	verify-with — every requirement carries a Verify with: line	hard error
C004	one-strength-word — exactly one of must / must not / should / should not / may	warning
C005	non-goals-present — a non-empty Non-goals section exists	warning
C006	open-questions-present — an Open questions section exists	warning
C007	no-tbd-at-ready — no TBD/TODO/unresolved question at status: ready	hard error
C008	sources-named — frontmatter sources: names at least one origin	warning
C009	broken-source-link — every named source resolves	hard error
C010	preserves-refs-resolve — change plan only: every preserved id resolves	hard error
C011	waves-present — change plan only: migration/rewrite/schema-change has waves	warning
C012	coverage — review packet only: in-scope ids covered, no orphan rows (non-draft spec)	warning
C013	verify-evidence-binding — review packet only: a structured verify block's cmd matches the requirement's named command + result=pass (non-draft spec)	warning
C014	do-not-change-touched — review packet only: a changed file matches a task's ## Do not change entry (distinct from out-of-scope drift)	warning
C015	citation-resolves — spec only: an inline [[KEY]] citation resolves to a matching <a id="KEY"> anchor in the named sources.md (skips when no sources.md is resolvable)	warning
C016	pass-needs-evidence — review packet (gate path): a Pass coverage row with an empty Evidence cell blocks (the reconcile path surfaces it advisorily)	hard error
C017	orphaned-reference — workspace: a bundled .agents/skills/<name>/references/<file> named nowhere in its SKILL.md (orphan direction only)	warning

Packet checks (checklist level): non-empty-paste — a completion claim binds to pasted output or a CI link; no-open-critical — nothing closes with an open blocking question; trigger-coverage — the Human attention section considered every exception class.

Appendix — reference values (producer note)

This appendix exists for producers of corpus tooling and documentation, not for adopters. The closed sets have exact sizes that tooling and fixtures reconcile against — the same eight rows, with their member lists, live in checks/README.md: block types (SOL form) 5 · strength words 5 · review results 7 (4 core + 3 lifecycle) · verification methods 9 · loop steps 6 (+ 2 conditional) · lifecycle steps (advanced) 9 · improve operations 10 · check layers 5 (S/P/M/V/O). These registry rows — the counts with their member lists — live in exactly two places: here and there. A change to any set updates both, and the fixtures, in one commit. (The numeral-bearing model names — the six-step loop, the nine-step lifecycle — are names, not registry copies.)

Related

• The basic workflow — the loop, per-shape flows, skip rules.
• Checks — every check in full, with the honesty legend.
• Glossary — every term on this page, defined.
• Step bars — what good output looks like, per step of the loop.
• Principles — the design rationale the rules on this page derive from.
• The advanced lifecycle — the finer-grained step model under the loop.