Spec Compound

spec-compound

Document a recently solved problem through role-based research; use parallel subagents only when dispatch is explicitly authorized and callable, otherwise run the same roles inline or serially.

Purpose

Captures problem solutions while context is fresh, creating structured documentation in docs/solutions/ with YAML frontmatter for searchability and future reference. Authorized dispatch can parallelize read-only research; correctness does not depend on it.

Why "compound"? Each documented solution compounds your team's knowledge. The first time you solve a problem takes research. Document it, and the next occurrence takes minutes. Knowledge compounds.

Workflow Contract Summary

• 输入： 一个最近解决且已有可回源验证的单一问题，或该问题带来的 durable project vocabulary。
• 输出： docs/solutions/ 下带 provenance、适用范围与失效条件的 learning，以及必要时对 CONCEPTS.md 的局部补充。
• 硬出口： 问题尚未解决、验证证据不足、一次请求包含多个独立 learning、目标 repo/source owner 不明确，或 promotion gate 不满足时不得写入 durable knowledge。
• 权威： 当前 source/test/log 和已验证 outcome 决定可沉淀事实；LLM 判断复用价值；只有 orchestrator 可写知识资产，dispatch 不授予 mutation。
• 消费者： 后续 spec-plan、spec-work、spec-debug、spec-code-review 与项目维护者。

Usage

spec-compound                            # Document the most recent fix
spec-compound [brief context]            # Provide additional context hint
spec-compound mode:headless              # Non-interactive run for automations
spec-compound mode:headless [context]    # Non-interactive run with context hint

One learning per run. The workflow's grounding, overlap detection, and cross-referencing all assume a single solved problem. When a session produced multiple distinct learnings, run the skill once per learning, sequentially — each run grounds fresh against the tree. Do not batch several learnings through one run and stitch cross-references between the drafts afterward; drafting-context numbering ("Learning 3") leaking into written docs is the failure this rule prevents.

CONCEPTS.md bootstrap requests

If invoked specifically to create or bootstrap CONCEPTS.md from scratch rather than to document a solved problem, do not run the normal phases — spec-compound populates CONCEPTS.md only as a side effect of documenting a real learning (it seeds the learning's area, not the whole repo; see Phase 2.4). Repo-wide concept-map creation is spec-compound-refresh's job. Redirect a standalone bootstrap request to spec-compound-refresh (which asks whether to build the concept map or run a refresh cycle), then exit.

Mode Detection

Check the invocation arguments supplied by the current host for the exact mode:headless token. Tokens starting with mode: are flags, not context — strip only recognized mode tokens while preserving the remainder, quoted paths, and token order before treating it as the brief context hint.

Headless mode is intended for automations and skill-to-skill invocation where no human is present to answer questions. The doc itself is identical to what an interactive Full run would produce — classification work (track, category, overlap) follows the same rules and writes nothing extra into the artifact. Once detected, headless mode applies for the entire run.

Pre-resolved context

Git branch (pre-resolved): !git rev-parse --abbrev-ref HEAD

If the line above resolved to a plain branch name (like feat/my-branch), use it in Phase 1 session-history filtering so the orchestrator does not waste a turn deriving it. If it still contains a backtick command string, shows an error, or is empty, derive the branch at runtime.

Repo root (pre-resolved): !git rev-parse --show-toplevel

If the line above resolved to an absolute path, use it as the session-history repo filter in Phase 1. If it still contains a backtick command string, shows an error, or is empty, derive the repo root at runtime with the shell tool (git rev-parse --show-toplevel, falling back to the working directory outside a git repo).

Support Files

These files are the durable contract for the workflow. Read them on-demand at the step that needs them — do not bulk-load at skill start.

• references/schema.yaml — canonical frontmatter fields and enum values (read when validating YAML)
• references/yaml-schema.md — category mapping from problem_type to directory (read when classifying)
• references/concepts-vocabulary.md — CONCEPTS.md format and inclusion rules (read in Phase 2.4 when domain terms surface)
• references/agents/session-historian.md — skill-local synthesis prompt for optional session-history compounding context (read only when explicit restricted-read authorization exists and the relevance gate escalates)
• references/grounding-validation.md — grounding-validation protocol: flag adjudication rules and the semantic validator prompt (read in Phase 2.45)
• assets/resolution-template.md — section structure for new docs (read when assembling)
• scripts/session-history/ — session discovery and extraction scripts copied into this skill so session-history support does not depend on the bundled session-history support
• scripts/validate-frontmatter.py — frontmatter parser-safety validator plus the opt-in --promotion exit gate for provenance/invalidation (run against the private candidate in Phase 2 step 6 through the existence guard documented there; resolves via the loaded skill directory anchor SKILL_DIR, with a manual-checklist fallback elsewhere)
• scripts/validate-doc-claims.py — mechanical claims validator: cited paths, commit SHAs, relative links, dangling drafting scaffold (run in Phase 2.45 via the SKILL_DIR anchor)

When spawning subagents, pass the relevant file contents into the task prompt so they have the contract without needing cross-skill paths.

Dispatch Authorization Boundary

在派发 repo profiler、research role、session-history synthesizer、semantic validator 或 specialized reviewer 前，记录：

worker_dispatch_authorization: authorized | missing
capability_probe: not_applicable | attempted | unavailable
worker_dispatch_capability: available | missing | unknown
worker_context_isolation: isolated | inherited | unknown
worker_model_override: supported | unsupported | unknown
worker_bounded_parallelism: supported | unsupported | unknown

workflow invocation does not authorize dispatch。Full/headless mode、上下文预算、scratch directory、权限设置或 prompt asset 存在都不构成授权。只有当前用户或可见 upstream handoff 明确请求 subagent、delegated work、persona 或 parallel work 时才可派发。缺授权时不得探测 tool schema，固定为 capability_probe: not_applicable + worker_dispatch_capability: unknown，依次 inline 或 serial 执行相同 role prompts 并记录 dispatch_authorization_missing。只有授权后才把 current-session registry/schema 作为 provider_untrusted evidence 检查：确认缺失时记录 subagent_capability_missing；surface 不可用、schema 不完整或候选不唯一时记录 worker_capability_unproven，均使用同一 fallback。隔离、模型覆盖和有界并发只取 live facts；required isolation 未满足时保持依赖 gate 打开，model unknown 时继承，parallelism unknown 时串行。记录 worker_dispatch_outcome。Fallback 保留 Context Analyzer、Solution Extractor、Related Docs Finder 等角色合同，但不得声称 independent subagent、fresh-context 或 parallel coverage。无论哪种路径，只有 orchestrator 可以写 docs/solutions/、CONCEPTS.md、instruction files 或任何 tracked path。

Execution Strategy

spec-compound does not ask the user which mode to run. Mode depends on context budget the agent can observe. Cross-session history is different: reading private session stores is a restricted-read boundary, so the workflow probes it only when the current user or visible upstream handoff explicitly authorizes that read; it never infers authorization from a compound request, Full mode, local file access, or tool availability. Missing authorization skips the probe with restricted_read_authorization_missing rather than opening another question. The only interactive prompt in the normal workflow is the Discoverability Check consent, because that one edits a tracked instruction file.

Mode selection (Full vs Lightweight) — decide it, don't ask it.

• Default to Full: the complete workflow (research, cross-referencing, overlap detection, grounding validation). This is the right choice for essentially every documented learning — its token cost is small next to the engineering work that produced the learning and is dwarfed by the value of a doc that compounds.
• Choose Lightweight (single-pass, no subagents — see Lightweight Mode) only when the learning is low-risk, bounded, source-grounded, and already backed by verification evidence, and either the session is near its context limit or the fix is trivial enough that cross-referencing would add nothing. Context pressure alone never waives promotion obligations. A learning is high-risk when a wrong or stale claim could materially weaken security/authorization, data integrity, migration/release safety, privacy, compliance, or irreversible mutation boundaries. High-risk learnings use Full mode; if the remaining context cannot support Full mode, leave a handoff or emit Documentation skipped instead of writing durable knowledge.
• State the chosen mode and a one-line reason as the first line of the completion output (e.g., "Ran Full mode." / "Ran Lightweight mode — session context was tight."). If Lightweight was the wrong call for the user's taste, re-running is a rare, cheap correction — cheaper than taxing every run with a prompt.

In headless mode, skip mode selection entirely and run Full Mode with session history disabled (Phase 1 step 4 omitted). Headless does not elevate dispatch authority; when the package-local boundary is not satisfied, proceed through the serial inline Full fallback.

Session history — an authorization-gated probe in Full mode. When explicit restricted-read authorization exists, Full mode runs the cheap discovery+metadata probe (Phase 1 step 4) and escalates to extraction+synthesis only when the probe surfaces genuinely relevant candidate sessions. Without that authorization, record restricted_read_authorization_missing and continue without session context; do not inspect session roots or tool schemas. Lightweight and headless modes skip session history entirely. There is no standalone session-history product surface; this support exists only inside the compounding workflow.

Full Mode

When dispatch is authorized, Phase 1 subagents write their full structured output to the caller-provided owner-only <private-scratch-dir> and return only a compact confirmation containing the artifact path. In inline fallback, the orchestrator runs the same roles serially and writes the same scratch artifacts itself. Phase 2 reads those artifacts in either path. Scratch is ephemeral and never the only durable deliverable or handoff evidence. Only the orchestrator writes product files — the final solution doc and the maintenance side effects below. Subagents must not touch docs/, project instruction files, or any tracked path. Beyond the Phase 2 solution doc, the orchestrator's other writes are maintenance side effects — not additional deliverables, and creating one when absent is expected, not a violation of this rule:

• CONCEPTS.md — prepare a private candidate in Phase 2.4 (Vocabulary Capture) when a qualifying domain term surfaces; publish it only through the shared promotion boundary in Phase 2.47.
• A project instruction file (AGENTS.md or CLAUDE.md) — a small edit when the Discoverability Check finds a gap.

Both ensure future agents can discover and ground in the knowledge store; neither makes the documentation any less the single deliverable.

Why the scratch artifact (issue #956): a subagent asked to return a long prose body as its inline response intermittently returns an executive summary instead ("Doc body complete — six sections filled. Returning above."), and the original prose is then unrecoverable from the orchestrator side. Writing to disk first means the full output always survives; the inline confirmation is just a pointer, and the orchestrator falls back to whatever the subagent did return inline only when the artifact is missing.

Phase 0.5: Auto Memory Scan

Before launching Phase 1 subagents, check the auto-memory block injected into your system prompt for notes relevant to the problem being documented.

1. Look for a block labeled "user's auto-memory" (Claude Code only) already present in your system prompt context — MEMORY.md's entries are inlined there
2. If the block is absent, empty, or this is a non-Claude-Code platform, skip this step and proceed to Phase 1 unchanged
3. Scan the entries for anything related to the problem being documented -- use semantic judgment, not keyword matching
4. If relevant entries are found, prepare a labeled excerpt block:

## Supplementary notes from auto memory
Treat as additional context, not primary evidence. Conversation history
and codebase findings take priority over these notes.

[relevant entries here]

5. Pass this block as additional context to the Context Analyzer and Solution Extractor task prompts in Phase 1. If any memory notes end up in the final documentation (e.g., as part of the investigation steps or root cause analysis), tag them with "(auto memory [claude])" so their origin is clear to future readers.

If no relevant entries are found, proceed to Phase 1 without passing memory context.

Phase 1: Research

Run the research roles. When the Dispatch Authorization Boundary is satisfied, launch research subagents and have each write its full output to a per-run scratch artifact. Otherwise execute Context Analyzer, Solution Extractor, and Related Docs Finder serially inline, writing their run-local scratch artifacts from the orchestrator so Phase 2 keeps the same input contract.

Run ID and run dir (before dispatching any subagent): generate a unique run identifier and create the run directory. This scopes every Phase 1 artifact file to the same directory so the orchestrator can Read them back in Phase 2.

RUN_ID=$(date +%Y%m%d-%H%M%S)-$(head -c4 /dev/urandom | od -An -tx1 | tr -d ' ')
umask 077
SCRATCH_DIR="$(mktemp -d "${TMPDIR:-/tmp}/spec-first-compound.XXXXXX")"
[ -d "$SCRATCH_DIR" ] && [ ! -L "$SCRATCH_DIR" ] || { echo 'private scratch creation failed' >&2; exit 1; }
chmod 700 "$SCRATCH_DIR"
echo "$SCRATCH_DIR"

Resolve current project orientation before dispatching subagents. Record the current target repo/worktree identity and dirty state when available, then read root instruction files and CONCEPTS.md directly for the vocabulary and conventions needed by the Context Analyzer. Keep this as run-local input with direct source refs; never persist or reuse it across runs, branches, or worktrees. If a source cannot be read, record that degraded fact and let the Context Analyzer limit its claims rather than substituting stale orientation.

Current source is the authority for code-behavior claims. Every promoted learning must retain direct source refs, observed revision/freshness, applicability scope, and an invalidation condition; session history, cached summaries, and external provider output are advisory leads only and cannot close grounding on their own.

CRITICAL — glob docs/solutions/ fresh every run. spec-compound writes new learnings there, so even a run-local orientation assembled earlier cannot stand in for the live enumeration in step 3.

Pass {run_id} and the verified <private-scratch-dir> into every Phase 1 subagent prompt. Recheck that the directory remains owned and non-symlink before publishing each file with same-directory temp + atomic rename. Each subagent writes its full structured output to its own file there, confirms the write succeeded (the file exists and is non-empty), and then returns only a one-line confirmation containing the artifact path — not the prose body inline. Artifact filenames by subagent:

• Context Analyzer → <private-scratch-dir>/context.json (frontmatter skeleton, category path, filename, track)
• Solution Extractor → <private-scratch-dir>/solution.md (the full doc-body prose sections)
• Related Docs Finder → <private-scratch-dir>/related.json (links, refresh candidates, overlap assessment)
• Session History synthesis subagent (when run) → <private-scratch-dir>/session-history.md (prose findings)

Return the full output inline whenever the artifact write did not succeed. This covers both cases where the orchestrator's Phase 2 inline fallback would otherwise have nothing to read: (a) {run_id} is empty or did not resolve (non-Claude-Code platforms where the pre-resolution failed), so there is no path to write to; and (b) {run_id} resolved but the write itself failed — tool permission denied, absolute-path writes unavailable, disk error, or the post-write existence check came back empty. In either case the subagent must return its complete structured output inline instead of a path, because the path would point at a file that does not exist. Return only the bare path when — and only when — the write is confirmed on disk. The artifact pattern is a reliability improvement, not a hard requirement; the orchestrator handles a missing artifact in Phase 2 by using the inline return.

Execution order:

• With authorized dispatch, launch Context Analyzer, Solution Extractor, and Related Docs Finder in bounded parallel. Without it, run the same roles serially inline and preserve their separate artifacts/results without presenting them as independent agents.
• Then, only when explicit restricted-read authorization exists, run the internal session-history discovery/extraction/synthesis flow (see step 4 below) in Full mode — skipped in lightweight and headless. Its cheap discovery+metadata probe runs after authorization and escalates only on a relevance hit. With separately authorized background dispatch it overlaps the research roles; in inline fallback it runs after the three serial research roles so one orchestrator does not interleave several context-heavy jobs. Without restricted-read authorization, record restricted_read_authorization_missing and do not inspect session roots or related tool schemas.

Research roles

1. Context Analyzer

• Extracts conversation history
• Reads references/schema.yaml for enum validation and track classification
• Determines the track (bug or knowledge) from the problem_type
• Identifies problem type, component, and track-appropriate fields:
  • Bug track: symptoms, root_cause, resolution_type
  • Knowledge track: applies_when (symptoms/root_cause/resolution_type optional)
• Incorporates auto memory excerpts (if provided by the orchestrator) as supplementary evidence
• Reads references/yaml-schema.md for category mapping into docs/solutions/
• Suggests a filename using the pattern [sanitized-problem-slug].md — no date suffix, even if existing files in the target directory have one; the date: frontmatter field is the canonical creation date
• Writes to context.json: YAML frontmatter skeleton (must include category: plus the promotion exit fields source_refs: and invalidation_condition:), category directory path, suggested filename, and which track applies. Returns only the artifact path.
• Does not invent enum values, categories, or frontmatter fields from memory; reads the schema and mapping files above
• Does not force bug-track fields onto knowledge-track learnings or vice versa

2. Solution Extractor

• Reads references/schema.yaml for track classification (bug vs knowledge)
• Adapts output structure based on the problem_type track
• Writes the full doc-body prose (all track-appropriate sections below) to solution.md and returns only the artifact path. This is the subagent most prone to the issue #956 summary-collapse, so its prose must land on disk rather than only in the inline return.
• Incorporates auto memory excerpts (if provided by the orchestrator) as supplementary evidence -- conversation history and the verified fix take priority; if memory notes contradict the conversation, note the contradiction as cautionary context
• Grounds code-behavior claims in source, not conversation memory. Before asserting how code behaves (enum values, status semantics, limits, defaults), Read the defining line at the current tree and cite file:line alongside the claim. A claim that cannot be verified against the tree is softened or attributed ("per this session's conclusion…"), never stated as fact
• Writes merge-state claims for time. Cite PR numbers rather than bare commit SHAs — SHAs are rewritten by rebase/squash merges and may not exist on other checkouts. A "fixed in X" claim requires the fix to be reachable from the current tree; otherwise phrase it as pending ("fix opened in #1608, unmerged as of this writing")

Bug track output sections:

• Problem: 1-2 sentence description of the issue
• Symptoms: Observable symptoms (error messages, behavior)
• What Didn't Work: Failed investigation attempts and why they failed
• Solution: The actual fix with code examples (before/after when applicable)
• Why This Works: Root cause explanation and why the solution addresses it
• Prevention: Strategies to avoid recurrence, best practices, and test cases. Include concrete code examples where applicable (e.g., gem configurations, test assertions, linting rules)

Knowledge track output sections:

• Context: What situation, gap, or friction prompted this guidance
• Guidance: The practice, pattern, or recommendation with code examples when useful
• Why This Matters: Rationale and impact of following or not following this guidance
• When to Apply: Conditions or situations where this applies
• Examples: Concrete before/after or usage examples showing the practice in action

3. Related Docs Finder

• Searches docs/solutions/ for related documentation
• Identifies cross-references and links
• Finds related GitHub issues
• Flags any related learning or pattern docs that may now be stale, contradicted, or overly broad
• Assesses overlap with the new doc being created across five dimensions: problem statement, root cause, solution approach, referenced files, and prevention rules. Score as:
  • High: 4-5 dimensions match — essentially the same problem solved again
  • Moderate: 2-3 dimensions match — same area but different angle or solution
  • Low: 0-1 dimensions match — related but distinct
• Writes to related.json: Links, relationships, refresh candidates, and overlap assessment (score + which dimensions matched). Returns only the artifact path.

Search strategy (grep-first filtering for efficiency):

1. Extract keywords from the problem context: module names, technical terms, error messages, component types
2. If the problem category is clear, narrow search to the matching docs/solutions/<category>/ directory
3. Use the native content-search tool (e.g., Grep in Claude Code) to pre-filter candidate files BEFORE reading any content. Run multiple searches in parallel, case-insensitive, targeting frontmatter fields. These are template patterns -- substitute actual keywords:
   • title:.*<keyword>
   • tags:.*(<keyword1>|<keyword2>)
   • module:.*<module name>
   • component:.*<component>
4. If search returns >25 candidates, re-run with more specific patterns. If <3, broaden to full content search
5. Read only frontmatter (first 30 lines) of candidate files to score relevance
6. Fully read only strong/moderate matches
7. Return distilled links and relationships, not raw file contents

GitHub issue search:

Prefer the gh CLI for searching related issues: gh issue list --search "<keywords>" --state all --limit 5. If gh is not installed, fall back to the GitHub MCP tools (e.g., unblocked data_retrieval) if available. If neither is available, skip GitHub issue search and note it was skipped in the output.

4. Session History (authorization-gated internal flow after the research block)

• Run only in Full mode with explicit restricted-read authorization. Without it, record restricted_read_authorization_missing, do not inspect session roots or related tool schemas, and continue to Phase 2 without session context. Skip entirely in lightweight mode or headless mode. After authorization, run a two-stage probe: the cheap discovery+metadata pass executes first, and the expensive extraction+synthesis executes only when the probe clears the relevance gate (see Escalation gate below).
• Run session discovery, branch/keyword filtering, scan-window selection, deep-dive selection, and per-session extraction directly inside this skill using scripts/session-history/.
• Read the skill-local synthesis prompt at references/agents/session-historian.md, then dispatch a generic subagent using that prompt content. Do not dispatch a standalone agent by type/name.

Session-history payload — keep tight. A long, keyword-rich payload licenses widening. Use this shape:

• Pre-resolved context (only if values resolved cleanly above; otherwise omit): repo name, current git branch.

• Time window: explicit 7 days unless the documented problem clearly spans a longer arc.

• Problem topic: one sentence naming the concrete issue — error message, module name, what broke and how it was fixed. Not a paragraph; not a bullet list of related topics.

• Filter rule (one line): "Only surface findings directly relevant to this specific problem. Ignore unrelated work from the same sessions or branches."

• Output schema:

  Structure your response with these sections (omit any with no findings):
  - What was tried before
  - What didn't work
  - Key decisions
  - Related context
  

Do not append additional context blocks, exclusion lists, or topic-keyword bullets — verbose payloads give the session-history flow license to keep widening the search and rapidly compound wall time. If keyword search is needed, the internal flow owns that decision based on the topic.

• Returns: structured digest of findings from prior sessions, or "no relevant prior sessions" if none found.
• Session history is the final Phase 1 input, not a workflow stop. When it returns, proceed directly to Phase 2 with its output as the last input — do not emit a summary and do not pause for the user. A "no relevant prior sessions" return is still a valid input; the documentation gets written without session context.

Script resolution. Set SKILL_DIR to the absolute path of the directory containing the SKILL.md you just read, and run the bundled scripts from "$SKILL_DIR/scripts/session-history/". Set SKILL_DIR inline in each bash block below (shell state does not persist between commands). If the bundled scripts are genuinely not present on disk under "$SKILL_DIR/scripts/session-history/", skip session history visibly with: "Session history bundled scripts were not found in this skill's directory; skipping the session-history probe for this run." Continue Phase 2 without session context.

Discovery pipeline. Infer the scan window from the problem topic, starting with 7 days. Run discovery and metadata extraction:

SKILL_DIR="<absolute path of the directory containing the SKILL.md you just read>"
if [ -f "$SKILL_DIR/scripts/session-history/discover-sessions.sh" ] && [ -f "$SKILL_DIR/scripts/session-history/extract-metadata.py" ]; then
  REPO_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
  REPO_NAME=$(basename "$REPO_ROOT")
  SCAN_DAYS="7"
  bash "$SKILL_DIR/scripts/session-history/discover-sessions.sh" "$REPO_NAME" "$SCAN_DAYS" --cwd "$REPO_ROOT" | tr '\n' '\0' | xargs -0 bash "$SKILL_DIR/scripts/run-python.sh" "$SKILL_DIR/scripts/session-history/extract-metadata.py" --cwd-filter "$REPO_ROOT"
else
  echo "Session history bundled scripts were not found in this skill's directory; skipping the session-history probe for this run."
fi

Pi sessions are included when present under ~/.pi/agent/sessions/; they carry cwd like Codex but no git branch. If _meta.files_processed is 0, return no relevant prior sessions. If the first pass finds no relevant branch matches, or if processing Codex or Pi sessions, derive 2-4 keywords from the topic and re-run metadata extraction with --keyword K1,K2,.... Keep at most 5 sessions across Claude Code, Codex, Cursor, and Pi, ranked by branch match, keyword match count, file size over 30KB, and recency. Exclude the current session.

Escalation gate. After restricted-read authorization, the discovery+metadata pass above is the cheap probe. Escalate to the extraction and synthesis stages below only when at least one retained candidate clears the relevance bar: a current-branch match, or ≥2 topic-keyword matches. If no candidate clears the bar (including the _meta.files_processed is 0 case), stop here, record no relevant prior sessions as the session-history input, and skip extraction and synthesis. This gate keeps the authorized probe cheap — the expensive synthesis is paid for only when a prior session is genuinely relevant.

Extraction pipeline. Create SCRATCH=$(mktemp -d -t spec-compound-sessions-XXXXXX). For each selected session, write extracted content to scratch files:

SKILL_DIR="<absolute path of the directory containing the SKILL.md you just read>"
if [ -f "$SKILL_DIR/scripts/session-history/extract-skeleton.py" ]; then
  bash "$SKILL_DIR/scripts/run-python.sh" "$SKILL_DIR/scripts/session-history/extract-skeleton.py" --output "$SCRATCH/<session-id>.skeleton.txt" < <session-file>
else
  echo "Session history bundled scripts were not found in this skill's directory; skipping the session-history probe for this run."
fi

Use extract-errors.py selectively when dead ends or recurring errors are likely useful. Pass only the scratch file paths and metadata to the synthesis subagent.

Synthesis dispatch. Build a generic subagent prompt containing:

• the full content of references/agents/session-historian.md
• problem_topic
• scratch_dir
• output_path: <private-scratch-dir>/session-history.md
• a sessions array with extracted file paths and metadata
• the output schema above
• the filter rule above

The subagent reads only the scratch paths, writes its prose findings to <private-scratch-dir>/session-history.md, and returns only that artifact path once the atomic write is confirmed. If {run_id} or the private scratch directory did not resolve, ownership/symlink recheck failed, or the artifact write failed, it returns the prose inline instead. If synthesis fails, note the failure and continue without session context.

Phase 2: Assembly & Candidate Validation

WAIT for all Phase 1 inputs to complete before proceeding — the three research roles (parallel only under authorized dispatch) and, when separately authorized in Full mode, the internal session-history flow, which may stop at no relevant prior sessions. An authorization skip is a terminal Phase 1 fact, not an empty permission to inspect private session roots.

The orchestrating agent (main conversation) performs these steps:

1. Collect Phase 1 results from the run artifacts. Read context.json, solution.md, related.json, and session-history.md when that flow ran. Under authorized dispatch, fall back to the subagent's inline return only when its artifact is absent or empty. Under inline fallback, the orchestrator owns both role execution and artifact writes. The artifact is authoritative when present.

2. Check the overlap assessment from the Related Docs Finder before deciding what to write:

   Overlap	Action
   High — existing doc covers the same problem, root cause, and solution	Update the existing doc with fresher context (new code examples, updated references, additional prevention tips) rather than creating a duplicate. The existing doc's path and structure stay the same.
   Moderate — same problem area but different angle, root cause, or solution	Create the new doc normally. Flag the overlap for Phase 2.5 to recommend consolidation review.
   Low or none	Create the new doc normally.

   The reason to update rather than create: two docs describing the same problem and solution will inevitably drift apart. The newer context is fresher and more trustworthy, so fold it into the existing doc rather than creating a second one that immediately needs consolidation.

   When updating an existing doc, preserve its file path and existing frontmatter structure, but add source_refs and invalidation_condition when absent because this path materially rewrites the learning. Update the solution, code examples, prevention tips, and any stale references. Add a last_updated: YYYY-MM-DD field to the frontmatter. Do not change the title unless the problem framing has materially shifted.

3. Incorporate session history findings (if available). When the internal session-history flow returned relevant prior-session context:

   • Fold investigation dead ends and failed approaches into the What Didn't Work section (bug track) or Context section (knowledge track)
   • Use cross-session patterns to enrich the Prevention or Why This Matters sections
   • Tag session-sourced content with "(session history)" so its origin is clear to future readers
   • If findings are thin or "no relevant prior sessions," proceed without session context

4. Assemble the complete markdown into <private-scratch-dir>/learning-candidate.md, reading assets/resolution-template.md for the section structure of new docs. Do not create or modify the final docs/solutions/** path yet. For an existing target, record its current existence and SHA-256 before assembly so publication can detect concurrent drift.

5. Validate the candidate frontmatter against references/schema.yaml, including non-empty source_refs and invalidation_condition promotion exit fields and the YAML-safety quoting rule for array items (see references/yaml-schema.md > YAML Safety Rules). The references must be grounded and the invalidation condition must be semantically specific; the script in step 6 checks only their mechanical shape.

6. Validate parser-safety and the knowledge-promotion exit contract on the candidate after every new or materially rewritten learning. Promotion mode catches malformed --- delimiter lines, unquoted # in scalar values (silent comment truncation), unquoted : in scalar values (silent mapping confusion), and mechanically requires a non-empty top-level source_refs array plus a non-empty top-level invalidation_condition. The bundled validator ships inside the skill bundle; SKILL_DIR resolves to the skill directory, but the runtime Bash tool's CWD is the user's project, so a project-relative path (without the $SKILL_DIR prefix) would miss. Run it through an existence guard so platforms that cannot locate the script (harnesses where $SKILL_DIR is unset) fall back to the same manual gate instead of silently skipping the protection:

   if [ -n "${SKILL_DIR:-}" ] && [ -f "$SKILL_DIR/scripts/validate-frontmatter.py" ]; then
     bash "$SKILL_DIR/scripts/run-python.sh" "$SKILL_DIR/scripts/validate-frontmatter.py" --promotion <candidate-path>
   else
     echo "Bundled validate-frontmatter.py not resolvable on this platform; applying the parser-safety and promotion checklist manually."
   fi
   

   • If the script ran: exit 0 means the mechanical promotion gate passed; exit 1 means stderr names the offending field(s) — repair the frontmatter and re-run until exit 0. Do not declare success while validation fails.
   • If the script did not run (else branch): apply the same parser-safety and promotion-shape checks by hand. Do not declare success until all four checks pass:
     1. The opening and closing frontmatter delimiters are each a line whose content is --- (trailing whitespace is fine; ---- or ---extra is not a valid delimiter).
     2. For each top-level mapping entry (key: value, no leading indentation) whose value is not already quoted or structured (does not start with ", ', [, {, |, or >): the value must contain no unquoted # (space-then-hash — YAML treats it as a comment and silently truncates) and no unquoted : (colon-then-space — strict YAML may read it as a nested mapping). Quote the whole value if either appears.
     3. source_refs appears exactly once as a top-level non-empty block or flow array, and every item is a non-empty string. Plain tokens that common YAML parsers type as null, boolean, number, sexagesimal, date, or timestamp do not count as strings; quote them.
     4. invalidation_condition appears exactly once as a top-level non-empty scalar or block string, with the same implicit-type quoting rule for plain scalar values. Nested parser-safety values, semantic source credibility, and semantic invalidation adequacy remain outside this mechanical fallback. Then state in the completion output that the bundled script validator was unavailable on this platform and the checks were applied manually.

   Default validator mode remains parser-safety-only for legacy compatibility. --promotion adds only the two promotion exit shapes; it does not judge reference credibility, invalidation adequacy, other schema fields, or enum values. It also does not flag YAML reserved-indicator characters (those produce loud parser errors downstream rather than silent corruption — out of scope). Uses Python 3 stdlib only (no PyYAML or other deps).

When creating a new doc, preserve the section order from assets/resolution-template.md unless the user explicitly asks for a different structure. A candidate passing this mechanical check is not promoted yet.

Phase 2.4: Vocabulary Capture

First, read references/concepts-vocabulary.md. This is unconditional. Do not pre-judge from memory that nothing qualifies — the reference's criteria are non-obvious and qualifying terms often live in the surrounding conversation rather than the new doc itself. Reading the reference is what makes the rest of the phase possible.

Then, applying those criteria, scan the learning candidate and the surrounding conversation for qualifying domain terms. Prepare any resulting CONCEPTS.md change as <private-scratch-dir>/concepts-candidate.md; do not modify the durable file before Phase 2.47. If CONCEPTS.md exists at repo root, base the candidate on its current contents and record its SHA-256; if it does not exist and at least one qualifying term surfaced, prepare a new candidate.

Verify behavior assertions against source before writing them. When an entry asserts how code behaves (states, transitions, limits, semantics

…(truncated)