docs-impact-classifier

Single responsibility: given a PR diff and the .apm/docs-index.yml corpus map, emit ONE classification verdict.

This skill is the cost gate for the entire docs-sync system. ~70% of PRs should exit at verdict no_change with zero panel spawn.

Architecture

This is a 3-layer funnel inside a single skill invocation:

• L0 deterministic path gate -- pure file-path matching, no LLM.
• L1 symbol extraction + corpus grep -- pure text processing, no LLM.
• L2 LLM classifier -- bounded ~8 KB context envelope, 1 call.

The skill returns the verdict from the earliest layer that can decide.

Step 1: L0 deterministic path gate (no LLM)

Read .apm/docs-index.yml to load no_impact_paths[] and user_surface_paths[]. Get the changed file list from the PR diff (gh pr diff --name-only).

if every changed file matches no_impact_paths AND none match user_surface_paths:
    return {verdict: "no_change", confidence: "high", source: "L0", scope_pages: []}

This handles:

• Test-only PRs (tests/**)
• CI workflow PRs (.github/workflows/**)
• Doc-only PRs (docs/**) -- out of scope, docs-sync doesn't review docs PRs
• Primitive-only PRs (.apm/**)
• Script and meta PRs

Expected hit rate: ~70% of PRs short-circuit here.

Step 2: L1 symbol extraction + corpus grep (no LLM)

If L0 did not exit, extract user-observable symbols from the diff:

• CLI command names -- grep diff for ^@click.command, ^@cli.command, or any apm <verb> mention in added/removed lines.
• Flag names -- grep diff for ^@click.option, --[a-z-]+ patterns.
• Public API symbols -- added/removed def <name> in src/apm_cli/__init__.py or src/apm_cli/api/**.
• Schema keys -- added/removed keys in apm.yml, apm.lock.yaml, apm-policy.yml parsers.
• Error strings -- added/removed string literals in user-facing error paths (look for _rich_error, click.echo, raise ... Error().

For each extracted symbol, consult .apm/docs-index.yml#symbol_index to find the documented pages. Collect all hits into candidate_pages[].

Also grep -rn <symbol> docs/src/content/docs/ for symbols NOT in the index (catches drift between index and corpus).

Step 3: L2 LLM verdict (1 call, bounded context)

If L1 found zero candidate pages AND zero schema/CLI/flag changes: return {verdict: "no_change", confidence: "medium", source: "L1", scope_pages: []}.

Otherwise, invoke the doc-analyser persona with EXACTLY this context envelope (must fit in ~8 KB tokens):

• PR title + body (first 500 chars)
• Diff stats (gh pr diff --stat output)
• .apm/docs-index.yml (the whole file; it's ~8 KB seeded, may grow)
• L1 candidate pages with +/-5 lines of context per hit
• Path-classification summary from L0
• pr_doc_diff_paths[]: the list of paths under docs/src/content/docs/** that the PR itself already modifies (drives the in_place_resolved downgrade rule in "In-place-resolved detection" below).

Ask doc-analyser to return JSON matching this schema:

{
  "verdict": "no_change" | "in_place_resolved" | "in_place" | "structural",
  "confidence": "low" | "medium" | "high",
  "scope_pages": ["docs/src/content/docs/..."],
  "structural_proposal": {
    "new_pages": [{"slug": "...", "rationale": "..."}],
    "moved_pages": [{"from": "...", "to": "..."}],
    "toc_changes": "<one-paragraph>"
  },
  "reasoning": "<one-paragraph: what surface changed, what docs are affected, why this verdict>"
}

structural_proposal is populated only when verdict is structural. scope_pages is populated for in_place and structural verdicts.

Verdict semantics

Verdict	Meaning	Panel size	Cost
no_change	No user-observable surface changed	0 panel spawns	~0-1 LLM call
in_place_resolved	Doc impact existed, but the PR's OWN diff already patches every page in scope_pages -- author already did the work	0 panel spawns; skill emits NO advisory	~1 LLM call
in_place	One to a few pages need a paragraph or section update; no new pages, no TOC change	N candidate pages x (doc-writer + python-architect) + editorial-owner + growth-hacker + CDO	~6-12 LLM calls
structural	A new page is needed, OR an existing page should be split/merged, OR the TOC needs to change to fit a new concept	architect first (TOC delta), then in-place panel for affected pages	~10-15 LLM calls

In-place-resolved detection (false-alarm killer)

BEFORE returning in_place, intersect your scope_pages[] with the list of files the PR itself touches under docs/** (provided to you by the orchestrator under pr_doc_diff_paths[]). If EVERY scope page already appears in pr_doc_diff_paths, downgrade to in_place_resolved and emit reasoning of the form "Author already patched <page list>". This is the well-behaved-author path; the skill stays silent.

If only SOME scope pages are pre-patched, keep in_place and list the REMAINING (unpatched) pages in scope_pages[]. Note the pre-patched ones in reasoning for transparency.

Rename / breaking-change heuristic (PR 1244 class)

When the L1 layer reports an ADDED public symbol that matches an EXISTING public symbol's name in the corpus (e.g. PR adds apm update but apm update already appears in 9 docs pages with different semantics), this is a RENAME or BREAKING SEMANTIC CHANGE. Bias toward structural (not in_place):

• the existing page describing the OLD semantics may need to SPLIT into two pages (old verb under new name + new verb keeping old name)
• the TOC may need a NEW reference page for the renamed verb
• every passing mention in the corpus needs verification

Do NOT collapse a rename into in_place just because the affected pages already exist. The shape of the work is structural even when no new page is strictly required.

Anti-patterns (verdict shape errors)

• Returning in_place with empty scope_pages -- invalid; orchestrator will reject.
• Returning structural without structural_proposal -- invalid.
• Returning in_place when EVERY scope page is in pr_doc_diff_paths -- should be in_place_resolved.
• Inflating structural to seem thorough -- the CDO will catch this. Return the minimal true verdict.
• Missing the rename heuristic above and emitting in_place for a verb-swap PR.
• Reading the corpus (the .md files themselves) at L2 -- context budget breach. You read the index, not the corpus.

Output contract

Return a SINGLE JSON document matching the schema in Step 3 as the final message of your task. No prose around the JSON. The orchestrator parses your last message.