Spec update

spec-update

Reconcile spec/*.md and CLAUDE.md against the current state of the code.

Modes

• No argument → audit all markdown specs, rewrite drifted ones, refresh INDEX + CLAUDE.md table. Pipeline-specs (HTML under spec/pipelines/) are not swept in this mode — they describe user-facing workflows, not code symbols, so the anchor-drift check would always treat them as unchecked and the agent has no signal to rewrite them. Update pipelines with the single-path form below.
• Single path / topic argument:
  • If the path matches spec/pipelines/<slug>.html → Read ${CLAUDE_PLUGIN_ROOT}/skills/spec-update/_pipeline-mode.md and follow it. Skip the rest of this file. Without a pipeline path, do NOT load that file.
  • Otherwise → update only that markdown spec, then refresh INDEX + table.

Preconditions

• spec/ directory exists and has at least one *.md. If not → tell the user to run /adam:setup and stop.
• CLAUDE.md exists with a spec index table.

Pre-step 1 — clean the gitnexus auto-injection

gitnexus analyze re-injects a <!-- gitnexus:start -->...<!-- gitnexus:end --> block into CLAUDE.md on every fresh run. Before delegating to the agent, run the strip script — idempotent and silent when the markers aren't present:

bash ${CLAUDE_PLUGIN_ROOT}/scripts/utils/strip-gitnexus-block.sh "$PWD"

This keeps the CLAUDE.md the agent reads in sync with what adam actually authors.

Pre-step 2 — anchor drift check (no-arg mode only)

When invoked without a path argument, run the anchor-drift script first to scope the agent's work to specs that actually need refreshing. Do not interpret these steps yourself; call the script.

bash ${CLAUDE_PLUGIN_ROOT}/scripts/tools/check-anchors.sh "$PWD"

The script walks every spec/**/*.md, parses the optional anchors: block list in frontmatter, asks GitNexus to resolve each entry, and emits JSON:

{
  "summary":   { "total", "drifted", "clean", "unchecked" },
  "drifted":   [ { "spec", "missing": [{ anchor, file, symbol, kind }] } ],
  "clean":     [ { "spec", "anchor_count" } ],
  "unchecked": [ { "spec", "reason" } ]
}

Hold the JSON. Then:

• drifted specs → known stale, dispatch the agent with a focused prompt naming the broken anchors (cheaper than a full re-read).
• clean specs → all anchors still resolve; skip them entirely.
• unchecked specs → no machine-readable anchors yet; fall back to the legacy "agent re-reads everything" path for these only.

If status != "ok" (no spec/, gitnexus missing, repo not indexed), surface the error and stop.

Process

Delegate to the adam sub-agent. The dispatch shape depends on what Pre-step 2 returned (no-arg mode) or on the user's path argument (single-spec mode).

No-arg mode, drift-driven:

Run a scoped drift refresh. The anchor checker has identified the following specs as drifted (anchors no longer resolve in the GitNexus graph):

<drifted block from check-anchors.sh>

For each drifted spec, open the code it describes, fix the broken anchors (the missing entries above are the seeds), and rewrite stale prose to match. Do not touch the specs in clean — their anchors all resolve, treat them as up-to-date. For specs in unchecked (no anchors: frontmatter yet), do a normal full re-read and add an anchors: block list to the frontmatter as you go so future runs can fast-path them. Refresh spec/INDEX.md and the CLAUDE.md spec table, re-counting tokens. Use the spec-lint MCP to verify the result.

Single-spec mode (user passed a path):

Update spec/<topic>.md against the current code at <paths>. Then refresh the row in spec/INDEX.md and CLAUDE.md (re-count tokens for that spec). If the spec already has an anchors: frontmatter block, refresh it to match the new code state.

(Pipeline-spec mode lives in _pipeline-mode.md — load it only when the path argument matches spec/pipelines/<slug>.html.)

Drift signals to look for

The agent should flag:

• File paths that no longer exist (e.g. spec mentions src/foo.py but code is at core/foo.py).
• Function / class / method names that have been renamed.
• Constants and thresholds that have changed value.
• DB tables / columns that have migrated.
• Stack claims that are wrong (e.g. spec says Node, code is Python).
• Dated "Current State (YYYY-MM-DD)" blocks more than ~3 months old that no longer match reality.

Output format

Show the agent's final report. The standard sections it produces — Rewrote / Deleted / Left alone / Lint / Notes — already cover what the user needs.

Guardrails

• Do not commit. Leave the working tree dirty.
• Do not touch TODO.md.
• Do not delete a spec just because the lint flagged a warning — only delete when the subject is truly gone from the code.
• Do not invent state. If a subsystem is unclear, leave a TODO: line in the spec and call it out in the notes section.