Docs Orchestrator

Docs Orchestrator Skill

Project-instruction file resolution: CLAUDE.md and AGENTS.md (Codex CLI) are transparent aliases — see skills/_shared/instruction-file-resolution.md.

docs-orchestrator coordinates the full documentation lifecycle inside a session: it detects which audiences (User, Dev, Vault) are touched by the agreed scope, generates audience-specific task definitions, threads them into the session-plan pipeline, and verifies that docs tasks produced diffs once waves complete. The skill is opt-in and default-off — when docs-orchestrator.enabled: false, all three hook points short-circuit with no output and no cost. docs-orchestrator fills the generative-content gap that sibling skills leave open: vault-sync validates but does not write, vault-mirror writes metrics-derived _overview.md entries but not narratives, claude-md-drift-check diagnoses CLAUDE.md (or AGENTS.md on Codex CLI) drift but does not remediate it, and daily exclusively owns 03-daily/*. docs-orchestrator is the only skill that produces new prose grounded in session output.

Invocation

Not user-invocable. Triggered at three hook points within the session lifecycle:

1. session-start Phase 2.5 "Docs Planning" — after user alignment, before handing off to session-plan. Reads the agreed scope, runs audience detection (Phase 2 below), and threads the detected audience list into the plan context so session-plan can classify tasks correctly.
2. session-plan Step 1.5 Agent Registry — docs-writer is added to the agent registry when docs tasks are present; tasks carrying role Docs are assigned to it (see session-plan Step 1.8).
3. session-end Phase 3.2 "Docs Verify" — after waves complete, verifies that each Docs-classified task produced a diff in the expected file-pattern target and reports gaps per docs-orchestrator.mode.

All three hook points are gated on docs-orchestrator.enabled: true. When disabled, every hook exits immediately after the config read.

Phase 0: Input Validation

When: At the start of every docs-orchestrator execution (all three hook points).

Action: Confirm the invocation context is valid before doing any work.

1. Caller check — Verify that this skill was invoked from one of the three recognised hook points: session-start Phase 2.5, session-plan Step 1.5, or session-end Phase 3.2. If the call context is missing or unrecognised, abort with:

   [docs-orchestrator] ERROR: Unexpected invocation context '<context>'. Expected one of:
   session-start Phase 2.5 | session-plan Step 1.5 | session-end Phase 3.2. Aborting.
   

2. Config gate — Read docs-orchestrator.enabled. If false, exit immediately with no output, no logging, no side effects.

3. Task spec validation (hook point 3 only — docs-writer dispatch tasks) — When wave-executor dispatches a Docs task to docs-writer, the task spec MUST contain:

   • audience ∈ {user, dev, vault} — reject any other value.
   • file-pattern — a non-empty glob matching a pattern from audience-mapping.md.
   • rationale — a non-empty string describing why this audience was triggered.

   If any field is missing or audience is not one of the three valid values, abort the task dispatch with:

   [docs-orchestrator] ERROR: Malformed task spec — missing or invalid field '<field>'.
   Required: audience ∈ {user,dev,vault}, file-pattern (non-empty glob), rationale
   (non-empty string). Aborting task dispatch.
   

   Do not fall through to Phase 1 with a malformed spec.

4. Mode validation — Confirm docs-orchestrator.mode ∈ {warn, strict, off}. Any other value is a config error; abort with:

   [docs-orchestrator] ERROR: Invalid mode '<value>'. Expected: warn | strict | off.
   

Phase 1: Read Session Config

Read Session Config per skills/_shared/config-reading.md. Extract:

• docs-orchestrator.enabled (boolean, default false)
• docs-orchestrator.audiences (list, default [user, dev, vault])
• docs-orchestrator.mode (warn | strict | off, default warn)

If enabled: false, exit immediately with no output. Do not log, do not query scope.

Config is read once at invocation; subsequent phases use the cached values. If the config block is absent entirely, all defaults apply and the skill proceeds as if enabled: false.

Phase 2: Audience Scope Detection

Given the agreed session scope (from the session-start Q&A), determine which audiences are touched by the planned work:

• User — new CLI flags or commands, breaking API changes, install-flow changes, new user-facing features, changed examples.
• Dev — architecture decisions, major refactors, new modules or subsystems, test coverage changes, dependency upgrades, ADR-level choices.
• Vault — project status changes, ownership transitions, stack or infra decisions, cross-project dependencies, migrations, archival events.

See audience-mapping.md (in this directory) for the authoritative file-pattern table, source rules per audience, and the non-overlap contracts with sibling skills.

Intersect the detected audiences with the docs-orchestrator.audiences config value. Subset selection is supported — e.g., audiences: [user, dev] omits Vault writing even when Vault signals are present in scope.

When mode is off: skip audience detection and return an empty list — no Docs tasks are generated.

Phase 3: Task Generation

For each selected audience, generate one or more docs tasks and thread them into the wave plan.

Each task MUST specify:

Tasks flow through the standard session-plan pipeline. They are role-classified as Docs and assigned to the docs-writer agent at Step 1.8. No custom dispatch path is required.

Audience routing — hard non-overlap rules:

Before finalising a task's file-pattern, check it against the forbidden targets:

• <vault>/01-projects/*/_overview.md — owned by vault-mirror. If the requested target matches this pattern, abort the task with:

  [docs-orchestrator] ERROR: Target '<path>' matches vault-mirror's owned pattern
  (<vault>/01-projects/*/_overview.md). Docs-writer must not write here. Aborting task.
  

• <vault>/03-daily/* — owned by daily. Same abort with the relevant pattern message.

If no valid non-forbidden target exists for a triggered audience, log an advisory (mode warn) or abort the session plan (mode strict):

[docs-orchestrator] WARN: Audience 'vault' triggered but all targets are forbidden.
No Docs task generated for this audience.

Phase 4: Source Grounding (docs-writer execution protocol)

This phase applies inside the docs-writer agent, not in the coordinator. It documents what the docs-writer MUST do when executing a dispatched Docs task.

Before writing any documentation, the docs-writer MUST read and ingest the actual source material provided in the task prompt. The canonical four source types are:

1. diff — git diff $SESSION_START_REF..HEAD — the verbatim diff of all changes made during this session. This is the primary authoritative source for what changed.
2. git-log — git log $SESSION_START_REF..HEAD --format="%H %s%n%b" — commit messages and associated bodies. Secondary context for all audiences; use to reconstruct the "why" when diffs alone are ambiguous.
3. session-memory — Session transcript, wave outputs, and agent summaries stored at ~/.claude/projects/<project>/memory/session-*.md and under .orchestrator/ for the current session. Primary source for Vault narratives (status updates, decisions made in conversation).
4. affected-files — Full content of the files listed in the wave-scope's allowedPaths (or the coordinator's affected-files context block). Primary for User and Dev content where understanding the updated interface, schema, or module structure is required.

No other sources are permitted. General knowledge about architectural patterns, Node.js APIs, or project history that is not present in the four sources above must not be used.

Source grounding steps:

1. Confirm all four source blocks are present in the task prompt. If a source block is empty or absent, note it before proceeding. Hard guard: if ALL four source blocks are empty or absent, ABORT this phase with error docs-writer: no grounding sources available — refusing to produce source-less documentation. Do not fall through to silent REVIEW-marker-only output — the coordinator must see the failure.
2. Read each source block in full before drafting any section.
3. For every paragraph drafted, identify which source (or sources) support it.
4. If a claim or narrative cannot be traced to at least one of the four sources, mark the paragraph with <!-- REVIEW: source needed --> inline — do NOT omit the marker to make output look cleaner.

Write or Update semantics:

• Prefer Edit on an existing file over Write (full replace).
• For multi-section updates, prefer surgical section-anchored Edits over full-file Writes.
• When the target file does not exist, use Write to create it.
• Respect existing document conventions: preserve heading levels, frontmatter structure, and table formats already present in the file.

Phase 5: Source Citations

This phase applies inside the docs-writer agent, concurrent with writing.

Every substantive new paragraph MUST include a source attribution in one of two forms:

Inline HTML comment (preferred for single-source paragraphs):

The coordinator now creates a git stash ref before dispatching each wave, enabling
crash recovery without data loss.
<!-- source: diff -->

Multi-source inline comment:

The new `--dry-run` flag skips file writes and prints a diff to stdout instead.
<!-- source: diff, affected-file:src/cli/index.ts -->

Collected ## Sources section (use when most paragraphs share the same sources):

## Sources

- diff: `git diff abc123..HEAD`
- git-log: commits `abc123`–`def456`
- session-memory: wave-2 coordinator summary
- affected-files: `src/cli/index.ts`, `SKILL.md`

Sourceless content:

If a section cannot be sourced from diff, git-log, session-memory, or affected-files, it MUST be marked:

<!-- REVIEW: source needed -->

This marker signals to the human reviewer that the content requires verification before the next release. The docs-writer MUST NOT remove this marker to make output appear complete.

Phase 6: Output Report

When: At the end of every docs-writer agent run, and aggregated by docs-orchestrator at session-end Phase 3.2 "Docs Verify".

The docs-writer MUST emit a brief structured output summary:

[docs-orchestrator] Docs task complete.
  Files touched: <comma-separated list of relative paths>
  Audiences served: <comma-separated list: user | dev | vault>
  REVIEW markers: <count> (files: <comma-separated paths or "none">)
  Source coverage: diff=<used|not used>, git-log=<used|not used>,
    session-memory=<used|not used>, affected-files=<used|not used>

Verification at session-end Phase 3.2:

1. Collect every task classified as Docs in the plan.
2. For each task, resolve the expected file-pattern target and check for a diff:

   git diff --name-only $SESSION_START_REF..HEAD
   

   Match the output against the task's file-pattern using glob matching.
3. Classify gaps by severity:
   • missing diff — no matching file was changed; the task did not run or produced no output.
   • partial — a matching diff exists but one or more sections contain <!-- REVIEW: source needed --> markers; needs human review before next release.
4. Report per docs-orchestrator.mode:
   • warn — log all gaps as advisories in the session final report. Non-blocking; /close proceeds normally.
   • strict — block /close until every gap is either resolved (diff exists, no REVIEW markers) or explicitly overridden by the user via AskUserQuestion.
   • off — skip verification entirely; emit no output.

Non-Overlap with Sibling Skills

• vault-sync — validates frontmatter schema; docs-orchestrator generates content. Complementary: vault-sync runs after docs-writer writes, catching any schema drift introduced by new docs.
• vault-mirror — writes _overview.md from JSONL metrics records; docs-orchestrator writes human-readable narratives in context.md, decisions.md, and people.md. Non-overlapping targets by design.
• claude-md-drift-check — detects drift in the project-instruction file (CLAUDE.md or AGENTS.md on Codex CLI; diagnostic only); docs-orchestrator remediates via the Dev audience path (generative). Complementary: drift-check can flag what docs-writer fixes. They must not run on the instruction file in parallel within the same session.
• daily — exclusively owns <vault>/03-daily/YYYY-MM-DD.md. This is a forbidden target for docs-orchestrator; the docs-writer prompt must carry this constraint explicitly (see Phase 3).

Session Config Reference

The three fields below are parsed in scripts/lib/config.mjs (_parseDocsOrchestrator) and validated in scripts/lib/config-schema.mjs (validateDocsOrchestrator). Defaults apply when the block is absent.

docs-orchestrator:
  enabled: false          # opt-in, default off
  audiences: [user, dev, vault]  # subset selection supported
  mode: warn              # warn | strict | off

• enabled — master switch. When false, all three hook points short-circuit.
• audiences — which audiences to generate tasks for. Any subset of [user, dev, vault] is valid.
• mode — verification strictness.
  • warn — surfaces gaps as advisories; /close is never blocked.
  • strict — blocks /close until all gaps are resolved or user-overridden.
  • off — disables audience detection, task generation, and verification entirely (lighter than enabled: false; config is still read but all execution paths exit early).

Anti-Patterns

• DO NOT write docs without a traceable source. Every section must trace to diff, git-log, session-memory, or affected-files, or carry <!-- REVIEW: source needed -->.
• DO NOT edit _overview.md or 03-daily/*. These are owned by vault-mirror and daily respectively. Attempting to write these targets aborts the task (Phase 3 hard non-overlap check).
• DO NOT duplicate the audience-mapping table. Always reference audience-mapping.md in this directory — never inline the table into a task prompt or another skill.
• DO NOT dispatch docs-writer outside the wave-executor flow. Docs tasks must go through the standard task → wave-executor → Agent() path so they appear in STATE.md, metrics, and the session final report.
• DO NOT use hard as a mode value. The canonical enum is warn | strict | off. Any other value fails config validation in Phase 0.
• DO NOT generate Docs tasks when mode is off. When mode is off, skip all execution after Phase 0 config read.

Source: hashgraph-online/awesome-codex-plugins → plugins/Kanevry/session-orchestrator/skills/docs-orchestrator/SKILL.md