Arch Docs

Arch Docs

Use this workflow only after explicit selection as $arch-docs or a binding task instruction requiring it. Preserve the repository's deletion and historical-evidence rules throughout the selected retirement work.

When to use

• The user wants a docs audit pass that finds stale docs, duplicate docs, outdated READMEs, dead migration notes, misleading setup/architecture docs, fake living docs, or grounded doc gaps that should exist in real evergreen docs.
• A repo needs topic-first cleanup where one topic may be spread across several files and several files may overlap on the same topic.
• A full-arch plan and worklog should be treated as context, mined for durable truth, and then retired or transformed into evergreen docs.
• The user wants to run the DGTFO loop directly in a repo with no requirement that an arch plan already exists.
• The user explicitly wants arch-docs auto and expects native goal-mode continuation with a new clean review after each pass.

When not to use

• The task is generic copy editing, README polish, or human-facing doc writing with no cleanup, clarification, canonicalization, or truth-audit workflow.
• The user wants net-new aspirational docs or a broad taxonomy redesign untethered from repo truth.
• The code is still changing too fast to ground docs honestly. Finish or audit the implementation first.
• The real task is choosing or editing a skill package. Use skill-authoring.

Non-negotiables

• The unit of work is the topic, not the file.
• Adapt to the repo's current doc gravity instead of imposing a new framework.
• Code is ground truth. Existing docs are evidence to inspect, not authority to trust.
• Resolve repo posture from repo evidence: public OSS versus private/internal.
• If repo posture is unclear, default to private/internal instead of guessing public OSS.
• Every in-scope topic should end with one canonical evergreen home.
• In public OSS repos, treat the standard community-doc surface as expected standalone canonical homes: README, LICENSE*, CONTRIBUTING.md, SECURITY.md, CODE_OF_CONDUCT.md, and SUPPORT.md.
• Bias toward deleting stale truth and promoting durable truth. Stale docs are worse than no docs, but missing or confusing canonical docs are also a failure.
• Standalone runs must still work with no DOC_PATH and no explicit path arguments.
• If explicit user context or active arch context exists, use it to narrow the audit intelligently. If not, audit the repo's docs surface directly.
• Use a temporary repo-root .doc-audit-ledger.md while the cleanup is active. Delete it before the cleanup finishes.
• Promote durable truth into evergreen docs, then delete obsolete working docs, plan/worklog residue, and stale duplicates in scope. Git is the archive.
• Do not trust what a doc calls itself. Folder names like docs/living and headers like Status: LIVING, Last verified, or similar freshness labels are claims to verify against code, not evidence that the doc should survive.
• Create a net-new doc only when it is the canonical result of the cleanup, not because the topic feels interesting.
• New docs are allowed in exactly two grounded cases:
  • a public OSS repo is missing a standard community-doc home
  • a differentiated evergreen topic fails the split-versus-expand test in references/canonical-home-judgment.md
• Deleting stale docs while leaving stale survivors, missing canonical truth, or confusing reader-critical docs behind is not a successful pass.
• Treat obviously time-bound docs as delete candidates by default once their durable truth has been folded forward. A doc does not survive just because it once mattered.
• In this repo family, point-in-time docs older than 30 days are presumed stale. Survival must be justified with explicit code-grounded current-reader value, not with doc labels, location, or freshness metadata.
• When a keep/delete call depends on whether a doc is stale, dated, or one-off, inspect git log and identify the last meaningful content change before deciding.
• Do not make a dead doc look current by editing freshness metadata alone. If the body was not materially re-grounded against current code in the same pass, the doc is still stale and should usually be deleted or folded forward.
• Before deleting any bounded batch of docs, stage those docs and create a backup git commit first. Stage only the docs in that delete batch, not unrelated dirty files elsewhere in the repo.
• Default behavior is one grounded docs-health pass. auto is the only explicit public mode.
• auto is goal-mode friendly. In native goal mode, keep running pass then review until the docs cleanup is clean or a real blocker stops it. Outside goal mode, run one bounded pass and name the exact next command.
• If the backup commit for a delete batch cannot be created, stop instead of deleting.
• If code truth is still unstable, external doc sources are required but unavailable, or the next pass would become speculative or materially unchanged, stop and say so plainly.
• Do not use age alone as the whole verdict. Use history and current code as evidence, but keep the 30-day presumption strong: old point-in-time docs should earn survival, not the other way around.
• Do not bury standard public-repo docs as README subsections when they should exist as their own conventional docs.
• Do not create docs just because a topic seems interesting. New or expanded docs must be grounded, canonical, and clearly useful to current readers.
• The suite's internal auto evaluator is allowed only when the ask explicitly says it is the internal evaluator. Do not suggest or advertise that internal surface to users.

First move

1. Read references/scope-and-profile.md.
2. Resolve the mode:
   • default one-pass DGTFO pass
   • auto only when the ask explicitly says auto
   • internal auto evaluator only when the ask explicitly says it is the suite's internal evaluator
3. Resolve working context:
   • explicit user paths, topic names, modules, or stop conditions
   • active arch plan, worklog, or clean implementation audit when present
   • otherwise the repo docs surface itself
4. Set LEDGER_PATH to .doc-audit-ledger.md and derive any scoped working artifacts only when they actually exist.
5. Read references/canonical-home-judgment.md.
6. Read references/cleanup-rules.md.
7. Read the mode reference:
   • references/pass.md
   • references/arch-docs-controller.md
   • references/internal-evaluator.md for the suite-only evaluator
8. Before dispatching any mapping or evaluator child, read ../_shared/agent-orchestration-policy.md.

Parent And Child Roles

• The active parent owns the resolved docs scope, ledger writes, child-result accounting, synthesis, accepted cleanup or repair decisions, and the final controller verdict. Capture current git status and the relevant diff before read-only child work, then compare current state before accepting the return.
• When a broad inventory benefits from mapping slices, and for the independent auto evaluator, default to new clean same-host native children. In Codex set fork_turns: "none"; in Claude use a clean named or custom subagent, not a bare conversation fork or skill context: fork shorthand. Use bounded or full inherited context only for a named dependency that exists solely in chat; ordinary scope travels through paths, the ledger, and a bounded brief.
• Give mapping and evaluator children the strongest read-only capability the host exposes and also tell them not to edit or write files, including the ledger. They may use their own native sub-agents on their own host; they may not start external agents.
• Split mapping only across independent topic or doc-surface families. Bound fanout by available host slots, shared-file or shared-state collision risk, and the parent's capacity to inspect and integrate every return.
• A background or external process used after the parent turn ends buys lifecycle continuity. It is not what makes an evaluator clean or independent; choose that transport only for its concrete lifecycle or other benefit under the shared policy.

Workflow

1) Default pass

• Profile the repo's doc system and resolve the strongest grounded scope.
• Inventory the relevant doc-shaped surfaces for that scope, using clean read-only mapping slices only when independent surface families make them worth the parent's synthesis cost.
• Ground each topic against current code and current shipped behavior.
• Treat every doc label, status line, and freshness header as untrusted until the code and current shipped behavior support it.
• Resolve repo posture from evidence and decide whether the public-repo baseline applies.
• Use git log when datedness or last meaningful use matters to a keep/delete call.
• For point-in-time docs older than 30 days, require an explicit code-grounded survival reason before they remain in scope as live docs.
• Collapse each topic to one canonical evergreen home, expanding existing homes first and creating a focused new one only when the canonical-home judgment says the topic should stand alone.
• Update stale surviving docs, clarify confusing docs, and promote missing grounded truth into the canonical home.
• Fold durable truth into an existing evergreen home whenever it fits cleanly. Most aged implementation-pass docs should disappear after that fold-forward instead of becoming standalone evergreen docs.
• Before deleting a bounded batch of stale, duplicate, dead, TEMP/WIP, obviously dated, or obsolete working-doc residue, stage those docs and make a backup commit, then delete them.
• Repair links, indexes, and navigation for the surviving canonical docs.
• Delete .doc-audit-ledger.md before finishing the run.

2) auto

auto is the repeated DGTFO loop. Native goal mode supplies the repeated turns; this skill does not install or arm automation hooks.

Workflow:

1. Run one grounded default DGTFO pass.
2. Run a new clean independent review/evaluator against the ledger and current code, using a same-host native critic by default while the parent session is active.
3. If review says more useful cleanup remains, run the next DGTFO pass.
4. In native goal mode, keep repeating until review says CLEAN or a real blocker stops the run.
5. Outside native goal mode, stop after one pass plus review and print the next exact command.

arch-docs-specific rules:

• User-facing invocation is arch-docs auto only when the ask explicitly says auto.
• Apply the same pre-delete backup-commit rule during each pass in auto.
• Keep sweeping while obviously dated docs, stale surviving docs, confusing docs, missing required public-repo docs, or missing grounded evergreen docs still remain.
• In this repo family, keep treating point-in-time docs older than 30 days as presumptively stale until the pass records explicit code-grounded current-reader value for each survivor.
• A pass that mainly refreshed labels like Status: LIVING or Last verified without materially re-grounding the body is still a stale-doc pass that needs more delete-first cleanup.
• Stop blocked when the review says the next pass would become speculative, taxonomy-imposing, disconnected from a narrowed scope, or materially unchanged.

Output expectations

• Keep console output short:
  • scope reminder
  • punchline
  • biggest docs-health risk or cleanup result
  • exact next move
• Put detailed inventory, topic mapping, deletion notes, and grounding evidence in .doc-audit-ledger.md while the cleanup is active.
• Do not leave the ledger behind when the cleanup is actually complete.

Reference map

• references/scope-and-profile.md - resolve repo profile, working scope, topic grouping, and canonical-home choices
• references/canonical-home-judgment.md - resolve repo posture, public-repo baseline docs, and split-versus-expand decisions for new docs
• references/cleanup-rules.md - deletion bias, canonical-home rules, working-doc retirement, and link-repair rules
• references/pass.md - the default DGTFO pass contract and ledger shape
• references/arch-docs-controller.md - arch-docs auto loop status and verdict source
• references/internal-evaluator.md - suite-only read-only evaluator contract