Forge playbook — Docdrift (external-doc claim drift detection)

Do not duplicate existing tools — this is a NEW, narrow capability. It does not replace forge-scout (external-capability research/vetting) or forge-audit-loop (this project's own internal health) — it only watches whether an external page a skill/doc CITES still backs up the claim made about it.

Provenance (honesty)

This tool's IDEA came from researching shanraisshan/claude-code-best-practice (an external community repo) — the Lead approved adopting the underlying idea ("keep external docs and our own claims about them in sync") as one of 3 GH-research winners. No code was copied. forge-bin/forge-docdrift.cjs is a Forge-native reimplementation built from scratch around this project's own conventions (zero-dep, config/orchestration/*.json seed + .claude/forge-research/* state/ledger, log-event.cjs's existing audit_finding event type, the same injectable-fetcher hermetic-test seam every other network-touching forge-bin tool uses). Provenance: PATTERN_ADAPTED (see the global research-recovery policy's honesty vocabulary) — the pattern was adapted, not the implementation.

What it does

forge-bin/forge-docdrift.cjs check [--source <id>] [--json] [--run <run_id>] reads config/orchestration/docdrift-sources.json — a seeded list of REAL claims this project's own skills/docs make that depend on an external, changeable page (an official Claude Code doc page, an n8n node doc page). For each claim it fetches the real source_url (following redirects, timeout-bounded, zero-dependency node:https/node:http) and searches the raw response body for every one of that claim's check_tokens.

Every claim is classified against its own persisted history in .claude/forge-research/docdrift-state.json:

• OK — tokens found, claim was not drifted last known check.
• NEW-DRIFT — a token is missing, claim was NOT drifted before (first time this claim looks stale).
• RECURRING — a token is missing again, same as last known check (a standing, already-reported gap).
• RESOLVED — tokens found again after a prior drift (the doc caught back up, or the claim was fixed).
• UNREACHABLE — the fetch failed/timed out/returned non-2xx. Never counts as drift — the last KNOWN drift state is left exactly as it was. A network hiccup is not proof the docs changed.

Every check appends one line to .claude/forge-research/docdrift-ledger.jsonl (append-only, never truncated) and persists the updated state. With --run <run_id>, ONE audit_finding event (already a registered log-event.cjs event type — no new registration needed) is logged per NEW-DRIFT result only — a RECURRING/RESOLVED/OK/UNREACHABLE result never re-fires an event for a condition already on record.

Honest limits (read before trusting a result)

• This is a heuristic token search, not a real diff against a known-good snapshot. A missing token can mean the docs genuinely changed — or it can mean an unrelated redesign/paraphrase changed the wording while the underlying behavior is unchanged. Every report says "possible drift, verify by hand" — never "the docs changed" as an assertion of fact.
• UNREACHABLE is not drift. A network failure, timeout, or non-2xx response tells you nothing about whether the claim still holds — it is reported honestly as unreachable, and it never overwrites the claim's last known drift state.
• Seeding a new claim requires reading BOTH files first: the citing skill/doc (to state the claim accurately) AND the live source page (to pick check_tokens that are genuinely present right now). Never invent a claim or a token set without checking both.
• Advisory only. A DRIFT/RECURRING result never blocks a build, a doctor run, or a WP handoff — it is a signal for a human (or a future Boss) to go verify the cited doc by hand.

When to use

• Before major n8n work (an n8n node's documented behavior — auth options, response mode — may have moved).
• Before relying on a cited Claude Code native feature (a subagent frontmatter field like memory/hooks/ isolation, or a slash command like /goal//loop) in a new skill or governance doc.
• On the owner's /forge docs-check.
• As one more check inside a periodic audit pass (alongside forge-audit-loop's own MEMORY-INTEGRITY/ AGENT-HEALTH/FEATURE-USAGE/DOCTOR-DELTA checks — this tool covers a gap none of those four checks touch: claims about content OUTSIDE this project).

Adding a new claim

Edit config/orchestration/docdrift-sources.json — add {id, skill, rule_id, category, check, check_tokens, source_url, added}. id must be unique; check_tokens must be a non-empty array of non-empty strings, each independently searched (all must be present for the claim to read as OK); source_url must be http(s)://. forge-docdrift.cjs's own config loader (loadSources()) refuses a malformed entry outright (missing field, empty check_tokens, non-http(s) URL, duplicate id) rather than silently skipping it.

Skills / commands

forge-bin/forge-docdrift.cjs (check — the only command; --source <id> to check one claim, --json for machine output, --run <run_id> to also log NEW-DRIFT findings as audit_finding events), config/orchestration/docdrift-sources.json (the seeded claim list — the single source of truth for what is checked), .claude/forge-research/docdrift-state.json (persisted drift memory, per-project, never a template file), .claude/forge-research/docdrift-ledger.jsonl (append-only audit trail, per-project).

Events

A real docdrift check with --run <run_id> logs audit_finding (the existing event type — never a new registration) once per NEW-DRIFT claim, carrying the claim id, rule id, missing tokens, and source URL as evidence. OK/RESOLVED/RECURRING/UNREACHABLE results are still recorded in the ledger but never generate a duplicate event for a condition already on record.

Fan-out & flow

1 agent, no team. Build Boss (or whichever Boss is doing the citing work) runs a docdrift check directly — this is a single zero-dependency CLI call, not a multi-agent task.

Ship-readiness (unique)

Every seeded claim's check_tokens were verified against the REAL live source page before being written to docdrift-sources.json (never invented); a real check was actually run (not merely described) before any report claims a verdict; every result names its literal status (OK/NEW-DRIFT/RECURRING/RESOLVED/UNREACHABLE) — never blurred into a single "drift found: yes/no"; UNREACHABLE is reported as UNREACHABLE, never silently treated as either OK or drift; nothing here ever blocks a build — it is advisory input for a human to verify.