update-agent-context

Update Agent Context

Overview

Keeps CLAUDE.md and AGENTS.md short, current, and auditable. CLAUDE.md is the source of truth; AGENTS.md is derived from it. The skill edits the working tree and shows a diff — it does not commit. Use /ship to commit after review.

Arguments: /update-agent-context [--phase1 | --phase2 | --phase3 | --all]

--all runs Phase 1, then Phase 2, then Phase 3 in sequence. Default when no flag is given: ask the user which phase to run.

Phase 3 is read-only — it audits the skill surface and produces a punch list. It does not edit any file. Run it after any watercooler-cloud tool-surface consolidation, or whenever a skill call unexpectedly fails because its tool name no longer exists.

---

Phase 1 — Structural Refactor or From-Scratch Seed

Run to (re)build CLAUDE.md on the approved structure and derive a fresh AGENTS.md. Phase 1 has two modes; Step 0 selects between them.

Step 0: Select mode

Check whether CLAUDE.md exists at the repo root.

• Refactor mode — CLAUDE.md exists. The current file is the primary content reservoir: carry its section content forward, fact-checked against source (Step 4). This is the common case (also use it after a major tool or role refactor).
• From-scratch seed mode — CLAUDE.md is absent. There is no reservoir; every section's content must be built from the grounding sources named in Steps 3–4 (source files, onboarding seeds, observed codebase patterns) plus the per-section content checklists in Step 5. Steps that read or diff the old file — Step 2, and the "carry forward" half of Step 4 — are skipped. Where a micro-convention has no file or seed source, infer it from the dominant pattern in the codebase and mark it inferred in the Step 8 diff so the user can confirm or correct it.

State the selected mode at the top of the run.

Step 1: Measure

Compute current sizes, measuring only files that exist (in from-scratch seed mode neither file is present, and a new repo may lack AGENTS.md even in refactor mode — a bare wc CLAUDE.md AGENTS.md would exit non-zero):

for f in CLAUDE.md AGENTS.md; do [ -f "$f" ] && wc -l -w "$f"; done

Record the numbers. The target for CLAUDE.md is ≤ 300 lines and ≤ 2,000 words (proxy for 2,000–3,000 tokens). Report both before and after. When a file does not exist its "before" is 0 — the loop above simply prints nothing for it.

Step 2: Record drift facts

Refactor mode only — skip in from-scratch seed mode (there is no prior file to drift from).

Read CLAUDE.md and AGENTS.md and note concrete mismatches that must be fixed, for example:

• AGENTS.md opens with file-based framing while CLAUDE.md says graph-first.
• Stale "stdlib-only" absolutes that conflict with current dependencies.
• Sections that duplicate role prose or tool inventory already covered by MCP.

Step 3: Fetch live role inventory (if Watercooler is available)

Load and call mcp__watercooler__watercooler_roles with code_path set to the repo root. Use the returned role names and summaries to confirm the six canonical roles (planner, critic, implementer, tester, pm, scribe). Do not copy role prose into CLAUDE.md; point to watercooler_roles instead.

If Watercooler is unavailable, proceed with local files only and note the limitation.

Step 4: Verify project facts against source

Before drafting prose, ground every concrete claim that will appear in the Project Snapshot, Repository Conventions, Development Workflow, and Security & Publishing sections. Read these source files directly and use only what they say. In refactor mode, do not carry a claim over from the old CLAUDE.md / AGENTS.md without re-checking it here; in from-scratch seed mode there is nothing to carry — build every claim directly from these sources.

Required reads:

• pyproject.toml — extract requires-python (the supported Python range), declared dependencies, and any deprecation notes. Do not state a wider support range than requires-python allows.

• .github/workflows/*.yml — extract the CI matrix (os: and python-version: lists). The "CI must pass on …" sentence in Development Workflow must name only the OSes that actually appear in the matrix.

• tests/ directory — list its top-level layout (tests/, tests/integration/, tests/unit/, etc.). Any "tests mirror sources" example must reference a file that actually exists. Prefer pointing at the directory pattern over a single example path so the bullet does not rot when files move.

• README.md "Requirements" / "Install" sections — cross-check Python version and install-extras claims.

• pyproject.toml [tool.black] / [tool.ruff] / [tool.mypy] tables (where present) — ground the Repository Conventions claims that are mechanically enforced: line length, import ordering (isort/ruff), and the type-check gate. Config-enforced conventions are facts, not preferences. Pared-down builds may omit these tables; when a table is absent, fall back to observed code and mark the item inferred.

• SECURITY.md — the vulnerability-disclosure contact and process; the single source for the Security & Publishing "Vulnerabilities" claim.

• A release / publish-pipeline config (a Copybara file, a release-automation workflow, etc.) if the repo has one — the source for any Security & Publishing "publish pipeline" claim. Many repos have none (and a downstream/open-core mirror typically does not ship its own publish config); skip silently when absent.

• CONTRIBUTING.md / .github/PULL_REQUEST_TEMPLATE* (if present) — branch and commit conventions and PR expectations.

• git log --format='%H %s%n%b' -30 — confirm the actual commit-message style (Conventional Commits or not) and whether commits carry a Signed-off-by trailer. The Development Workflow commit rules must match observed history, not an aspiration. The sign-off identity itself is not invented here — if a sign-off convention is in use, point at it; do not hardcode an address.

• Watercooler onboarding seeds (when present) — discover seed threads written by the watercooler-onboarding skill via the onboarding thread tag. These are higher-trust than the prior CLAUDE.md because their bodies cite source files with path:line-range provenance and were curated specifically as durable foundation material.

  Discovery (preferred — tag-based):

  watercooler_list_threads(tags="onboarding", code_path=<repo_root>)
  

  Returns every onboarding-tagged thread without requiring you to know the canonical topic names. New canonical topics added by future onboarding releases are picked up automatically.

  Discovery (fallback — when no threads carry the tag): Probe the canonical topic list directly. This is needed for repos seeded before the tagging convention landed (2026-05-04). After the first successful read, surface in the report that the seeds are untagged and recommend a watercooler-onboarding refresh run to apply the tag.

  Read the most recent entry of each topic that exists:

  • onboarding-architecture — subsystem boundaries + active architectural constraints. Use to ground Project Snapshot and the Watercooler Protocol section (graph-first reads, tier model, transport).
  • onboarding-working-map — directory map and entrypoints. Use to ground Project Snapshot layout claims.
  • onboarding-team-map — CODEOWNERS-derived ownership + recent contributors. Use when Repository Conventions mentions ownership / sign-off conventions.
  • onboarding-risk-register — volatile paths and drift findings. Use to surface cautions an agent should know before changing the system; record any drift the seed flags as candidates for fixing in this same patch.
  • onboarding-docs-contracts — synchronized doc / API / schema surfaces. Use to ground claims about what stays in sync (e.g. CHANGELOG, manifests, OpenAPI).
  • onboarding-entry-path — recommended first tasks per role. Cross-check against Development Workflow wording.
  • onboarding-overview / onboarding-product-charter — front-door framing and product bet. Use to ground the one-line product description in Project Snapshot.

  Canonical topic names (for fallback probing only — tag-based discovery is preferred): onboarding-overview, onboarding-product-charter, onboarding-team-map, onboarding-architecture, onboarding-working-map, onboarding-risk-register, onboarding-test-surface, onboarding-docs-contracts, onboarding-entry-path. Optional: onboarding-developer-experience, onboarding-release-process, onboarding-security, onboarding-recent-activity. The slug prefix produces a default thread title beginning with "onboarding " in the dashboard. Onboarding-written entries also start with the entry-title prefix Onboarding: (or Onboarding refresh: for refresh runs) — useful for distinguishing seed entries from ad-hoc additions in the same thread.

  Legacy / pre-2026-05-04 seeds may exist under bare slugs (overview, architecture, etc.). When the prefixed-slug fallback returns empty, fall through to the bare-slug list as a last resort and recommend a /watercooler-onboarding refresh to migrate.

  Fetch each discovered thread with watercooler_read_thread(topic=<name>, code_path=<repo_root>, code_branch="*"). If neither tag-based discovery nor canonical-topic probing returns any seeds, none exist — proceed with local files only and note in Step 2's drift facts that running /watercooler-onboarding would strengthen the next refresh.

  Treat seed entries as snapshots in time: cross-check named symbols / files / flags against the current source the same way Phase 2 Step 7 validates existing convention bullets.

If any existing claim in the current CLAUDE.md / AGENTS.md contradicts these sources, treat the source files as authoritative and rewrite the claim. Record each correction in the drift-fact list from Step 2 so it is visible in the diff.

Step 5: Rewrite CLAUDE.md

Apply the approved section structure below. Write the full replacement in one pass.

Target structure:

## Project Snapshot          (~150 tokens)
## Behavioral Principles     (~350–450 tokens)
## Repository Conventions    (~500–700 tokens)
## Development Workflow      (~250–350 tokens)
## Watercooler Protocol      (~400–600 tokens)
## Security & Publishing     (~150–250 tokens)
## Project Conventions       (~300–500 tokens; generated block — see Phase 2)
<!-- agent-context-refresh:last-updated=YYYY-MM-DD -->

Behavioral Principles — include all four Karpathy principles with short project-local adaptations. Do not copy a large external block; name each principle and add one sentence of local context:

1. Think Before Coding — State assumptions explicitly. Surface tradeoffs. If multiple interpretations exist, present them; do not pick silently.
2. Simplicity First — Minimum code that solves the stated problem. No speculative features, no abstractions for single-use code.
3. Surgical Changes — Touch only what you must. Match existing style. Do not improve adjacent code unless it is directly in scope.
4. Goal-Driven Execution — Transform tasks into verifiable goals. State a brief plan for multi-step work; loop until verified.

Project Snapshot — a ~150-token orientation: one-line product description, supported language/runtime range, core tooling, test layout, and the docs vs internal-docs split. Ground the facts per Step 4; the product one-liner comes from onboarding-overview / onboarding-product-charter (or README.md when no seed exists). The docs/ (user-facing) vs internal-docs split and any solution-writeup location are conventions — source them from an onboarding seed or a Decision entry, not invention.

Repository Conventions — enumerate the project's concrete conventions, one short bullet each. Cover this checklist; source each item as noted, and in from-scratch mode mark any item with no file/seed source as inferred:

• Layout — package layout, test-directory layout, bundled-data location. Source: observed directory structure (Step 4 tests/ read + a top-level ls).
• Type hints / type-check gate — whether hints are required and the checker. Source: [tool.mypy] config + observed signatures.
• Docstrings — style (e.g. Google) and where required. Source: onboarding seed or the dominant codebase pattern.
• Errors — exception-raising norms. Source: seed or dominant pattern.
• Imports — ordering, absolute vs relative. Source: [tool.ruff] isort config (config-enforced = fact, not preference).
• Naming — casing per identifier kind, CLI command casing, role names, entry-type names. Source: role names from watercooler_roles; entry types from the code enum; casing from observed code.
• Dependencies — core-minimal policy and the bar for adding a dependency. Source: a Decision entry or onboarding seed.
• Comments — default density. Source: Behavioral Principles + dominant pattern.
• Coverage — target, if any. Source: CI config or a Decision entry; mark inferred if neither states a number.
• Markdown — wrap width and heading style. Source: [tool.black] / editorconfig line length where it applies; otherwise observed.

Development Workflow — cover this checklist, grounded per Step 4:

• Branches — naming / prefix scheme. Source: CONTRIBUTING.md or observed branch names in git log / git branch -a.
• Commits — message convention and any Signed-off-by requirement. Source: the git log read in Step 4 — state only what history actually shows.
• Pull requests — required CI and merge style. Source: .github/workflows/ for the CI matrix (name only OSes/versions that appear); merge style from CONTRIBUTING.md or observed merge commits.
• Tests — how to run them and notable markers. Source: pyproject.toml pytest config + tests/ layout.
• Quality gate — the exact lint/format/type command(s). Source: the [tool.*] configs + any Makefile / CI lint step.
• Install — the dev-install command and required extras. Source: pyproject.toml extras + README.md.
• Risky actions — the confirm-before rule for destructive / shared-state operations. Source: Behavioral Principles (carry the project-local wording).

Watercooler Protocol — condense to a single focused section. Must preserve:

• Identity setup: per-call agent_func in the format <platform>:<model>:<role>. There is no set_agent / watercooler_set_agent shortcut — set agent_func on every write call (write, say, ack, handoff, set_status). Verify against the registered tool surface in src/watercooler_mcp/capabilities.py before claiming a setter tool exists.
• code_path discipline: always set to the repo root being worked on.
• Graph-first reads: baseline graph (JSON) is the sole source of truth; .md files are write-only projections. Always start with summary_only=true.
• Mandatory MCP use: never read/write thread files directly.
• Write discipline: watercooler_write is the preferred first-choice write tool (it wraps say/ack/handoff); the primitives say / ack / handoff / set_status remain available for explicit coordination control. All writes go through MCP tools only.
• Spec: marker: first line of every entry body.
• Pointer to watercooler_roles for full role behavior — do not repeat prose.
• Intent-first modification rule: before modifying existing code or state, query watercooler threads and semantic search to understand the intent behind the current state. Do not infer design constraints from code alone.
• Proactive recall: run watercooler_smart_query before starting significant work.

Security & Publishing — cover this checklist, grounded per Step 4. Omit any bullet whose subject the project does not have — the checklist is a menu, not a mandate:

• Secrets — the never-commit rule and where required env vars are documented.
• Input validation — the validate-at-boundaries norm. Source: Behavioral Principles or a Decision entry.
• Vulnerabilities — the disclosure contact and process, taken from SECURITY.md verbatim. Do not invent an address.
• Reference-only code surfaces — when the project has a deployed-vs-reference split (e.g. a Slack formatter whose production copy lives elsewhere), keep a one-line warning naming which file is authoritative. Load-bearing: it stops agents editing the wrong file.
• Publish pipeline — include only when the repo publishes a downstream artifact or mirror: the pipeline and the rule against hand-pushing it. Source: the release/publish-pipeline config found in Step 4. A repo that is the published downstream (e.g. an open-core mirror) usually has no such config — omit the bullet.

Sections to eliminate or replace with pointers:

• Full docstring / type-hint examples → keep as one-line bullet, remove code blocks
• Testing guidelines with code blocks → keep as bullets only
• Module organization inventory → remove (derivable from codebase)
• Role definitions section → replace with pointer to watercooler_roles
• Full tool usage protocol → replace with condensed Watercooler Protocol section
• Verbose Slack integration architecture prose → cut. But keep the one-line reference-only warning (per the Security & Publishing checklist) — that warning must survive the refactor, not be removed with the prose around it.

Step 6: Validate invariants

Before writing, verify the draft preserves all of these. If any is missing, add it.

• Graph-first reads stated; .md files are projections, not source of truth
• Mandatory MCP use for all thread / watercooler operations
• Identity setup names per-call agent_func only — no set_agent / watercooler_set_agent reference (verify against src/watercooler_mcp/capabilities.py)
• Spec: marker write discipline preserved
• Write discipline names watercooler_write as the preferred write tool, with say / ack / handoff / set_status as the explicit-control fallback
• code_path alignment rule
• Intent-first modification rule (query threads before changing state)
• No repeated role prose (pointer to watercooler_roles only)
• No large tool inventories
• No stale "stdlib-only" absolutes that conflict with current deps
• Project Snapshot facts match source: Python range matches pyproject.toml:requires-python, OS list matches .github/workflows/, test-layout examples reference paths that actually exist
• ## Project Conventions block present (may be empty; Phase 2 fills it)
• Footer comment: <!-- agent-context-refresh:last-updated=YYYY-MM-DD -->

Step 7: Derive AGENTS.md

AGENTS.md is CLAUDE.md with the Claude-Code-specific mechanics removed — it keeps every section, including ## Watercooler Protocol. Apply these edits to a copy of the new CLAUDE.md:

• Drop the **Tool invocation:** subsection of ## Watercooler Protocol — the ToolSearch / mcp__<server>__<tool> / Serena call mechanics are Claude-Code-specific. Keep the rest of the section (graph-first reads, intent-first rule, identity, write discipline, storage contract).
• Neutralize Claude-Code-flavored examples — agent_func examples and the like should use platform-neutral identities, not a single client's.
• Drop skill-name references — replace <skill-name> skill mentions (e.g. "follow the editorial-publishing skill") with the underlying doc or process; AGENTS.md is read by agents that have no Claude Code skill catalog.
• Cut anything referencing Claude Code hooks or MCP-client-specific identity details.

Do not strip whole sections by header name — the difference between the two files is small and subsection-level. Preserve everything else verbatim in intent: repo facts, graph-first claims, security rules, commit policy (including the sign-off rule), Karpathy behavioral principles, mandatory Watercooler MCP requirements, intent-first rule, and the full ## Project Conventions block.

Add this marker as the first line of AGENTS.md:

<!-- generated-from: CLAUDE.md by /update-agent-context -->

Step 8: Check budget and show diff

Re-run wc -l -w CLAUDE.md AGENTS.md. If CLAUDE.md exceeds 300 lines or 2,000 words, trim verbose prose — prefer bullets over paragraphs. Then show the diff and stop. The user reviews and commits via /ship.

---

Phase 2 — Convention Extraction

Run periodically (monthly or after a significant batch of decisions) to patch the ## Project Conventions generated block with durable conventions from Watercooler memory.

Step 1: Determine the extraction window

Read the footer comment from CLAUDE.md:

<!-- agent-context-refresh:last-updated=YYYY-MM-DD -->

If absent, default to the last 30 days and state that this is a bootstrap run.

Step 2: Fetch Decision entries (T1 — always attempt)

Load and call:

mcp__watercooler__watercooler_list_decisions(
    since=<last_update_date>,
    limit=500,
    code_path=<repo_root>
)

If Watercooler is unavailable, stop Phase 2 with a clear message. Convention extraction without evidence would be unauditable.

Step 3: Broad entry search (T1 — always attempt)

Load and call:

mcp__watercooler__watercooler_search(
    mode="entries",
    query="convention naming workflow must never always rule pattern",
    query_operator="OR",
    start_time=<last_update_iso>,
    code_path=<repo_root>
)

Step 4: Optional enrichment (T2 — catch errors and continue)

Load and call:

mcp__watercooler__watercooler_smart_query(
    query="What durable project conventions were established or changed since <date>?",
    max_tiers=2,
    resolve_provenance=True,
    code_path=<repo_root>
)

If this call errors (open-core T1-only install, network issue, etc.), continue with T1 evidence only. Do not gate on watercooler_health — call directly and inspect the result.

Step 4.5: Pull onboarding seed entries (when present)

The watercooler-onboarding skill writes a canonical set of seed threads with strict body templates (Spec / Purpose / Observed / Inferred / Drift findings / Provenance). Their Observed: "active architectural constraints" items, Inferred: high-confidence claims, and Drift findings: rules are high-quality convention material — curated, file:line-cited, human-authored, and explicitly designed as durable foundation knowledge.

Discovery (preferred — tag-based):

watercooler_list_threads(tags="onboarding", code_path=<repo_root>)

Filter the convention-relevant subset (onboarding-architecture, onboarding-risk-register, onboarding-docs-contracts, onboarding-team-map, onboarding-entry-path) from the returned topic list, then read each:

mcp__watercooler__watercooler_read_thread(
    topic=<name>, code_path=<repo_root>, code_branch="*"
)

Discovery (fallback — when tag query returns empty): probe the canonical topic names directly. Repos seeded before the onboarding-tagging convention landed (2026-05-04) won't have the tag yet. After successfully reading any untagged seed, surface in the candidate report that a /watercooler-onboarding refresh would apply the tag and recommend running it.

For each topic that returns content, scan the body for rule-like passages in two sections:

(a) Observed: and Drift findings: — primary scan. This is where forward rules and sync obligations are most often stated outright.

• architecture — items under "Active architectural constraints" or "Load-bearing invariants" are convention candidates.
• risk-register — Drift findings: items framed as "X must stay in sync with Y" are convention candidates; one-off operational risks are not.
• docs-contracts — synchronization rules (e.g. "contract-affecting changes must update docs and generated references in the same PR") are convention candidates.
• team-map — CODEOWNERS-derived ownership rules are convention candidates only when they encode an enforceable workflow (e.g. mandatory reviewers).
• entry-path — recommended workflows per role are convention candidates when they apply to all PRs in their scope.

(b) Inferred: — secondary scan, filtered. Most inferred items are framing or positioning prose ("repo is built for X") that does not translate to a forward rule and should be skipped. A subset, however, is genuine convention material. Promote an inferred item to a candidate only when all four of these hold:

1. confidence: high (medium / low never qualify);
2. basis: cites a file, line range, or code surface — not a vibe;
3. The claim is framed as a behavioral norm or a framing-trap warning, not as descriptive prose. Recognizable shapes:
   • "X is the authority for Y" (where to look for ownership / source-of-truth)
   • "Conflating A and B is a [known] trap" (axis-orthogonality warnings)
   • "Treat C as D, not E" (interpretive norms)
   • "Don't infer X from Y" (anti-pattern guidance)
4. The norm is durable and cross-cutting — it would still apply six months from now and across more than one narrow file.

Worked example (2026-05-04): onboarding-team-map Inferred item "the repo treats CODEOWNERS as 'default reviewer signal,' not 'per-path accountability' — this seed is the canonical authority for ownership questions, not CODEOWNERS itself" is a candidate (high confidence, basis cites .github/CODEOWNERS:1-2 + git shortlog, framed as "X is the authority for Y", durable until the team grows past single-glob CODEOWNERS). Counter-example: onboarding-architecture Inferred item "the repo is structured for a long-lived multi-tier memory product, not a small CLI" is descriptive positioning — skip.

(c) Inferred: — risk-signal scan (parallel output, not convention material). Items that fail the section (b) filter may still carry useful "watch this" or "monitor that" value. These do not belong in CLAUDE.md's Project Conventions block (which is for forward behavioral rules) but should be surfaced in the Step 6 candidate report as "Risk signals to forward to next risk-register refresh" so the user can decide whether to fold them into the next /watercooler-onboarding refresh of onboarding-risk-register.

Promote an inferred item to a risk signal when all three hold:

1. confidence: medium or confidence: high (low never qualifies);
2. basis: cites a measurable surface — a number, a counter, an observable ratio, a recurring pattern with at least one concrete instance — not pure speculation;
3. The claim is framed as monitoring guidance or a recurring-pattern warning, not as descriptive prose. Recognizable shapes:
   • "X suggests Y" / "X indicates Y" (interpretive monitoring claim)
   • "Watch for X" / "X is worth monitoring" (explicit monitoring call)
   • "X is a recurring pattern" / "this class of drift" (pattern escalation candidate — codify a forward convention if observed again)
   • Numeric thresholds without a current rule ("ratio of N% suggests ..." where N has no codified band).

Risk signals are reported informationally only — they never auto-patch CLAUDE.md. The output is a list the user reviews and optionally forwards to a risk-register refresh.

Worked examples (2026-05-04 — three items today's Phase 2 would have surfaced if this scan had been in place):

• DLQ ratio. onboarding-risk-register Inferred item "The dead-letter ratio (~37%) suggests the memory pipeline has been working through a real backlog" — medium confidence, basis cites watercooler_health counters, framed as "X suggests Y". Forward to risk-register; consider noting a healthy DLQ band.
• CHANGELOG lag as recurring pattern. onboarding-docs-contracts Inferred item "the team ships releases via main → staging → stable → tag but CHANGELOG updates don't ride along automatically" — medium confidence, basis cites [0.4.0] only documented vs three shipped tags, framed as recurring pattern. Forward to risk-register; if observed twice more post-fix, codify as a forward convention.
• Self-description framing drift as a class. The specific "File-based" instance from onboarding-risk-register is being fixed in PR #757, but the class — package docstring / README / thesis must agree on framing — is a recurring drift surface. Forward to risk-register as a monitoring item for major doc refreshes.

Skip overview, product-charter, working-map, test-surface for convention extraction — they are framing or navigational, not rule-establishing.

If no canonical onboarding topics exist, skip this step. Note in the candidate report that /watercooler-onboarding has not run for this repo; running it would expose another high-trust convention source on the next refresh.

Step 5: Pull daemon advisory findings

The decision-stance, pulse-snapshot, and project-coordinator daemons emit findings that surface convention drift, stalled threads, role-imbalance signals, and coordinator-led improvement opportunities. These are complementary to the keyword/semantic search in Steps 2–4 and feed both candidate scoring (Step 6) and existing-bullet re-validation (Step 7).

Call all three; non-running daemons return empty results, so calling speculatively is safe:

mcp__watercooler__watercooler_daemon_findings(
    daemon="project_coordinator", limit=30, unacknowledged_only=True,
    enrich=True, code_path=<repo_root>
)
mcp__watercooler__watercooler_daemon_findings(
    daemon="decision_stance", category="stance_advisory", limit=20,
    unacknowledged_only=True, code_path=<repo_root>
)
mcp__watercooler__watercooler_daemon_findings(
    daemon="pulse_snapshot", enrich=True, limit=20, code_path=<repo_root>
)

Notes on these calls:

• project_coordinator and decision_stance are mutually exclusive — only one is registered at a time (daemons/__init__.py deregisters decision_stance when the coordinator is active). Whichever is unregistered returns an empty list.
• enrich=True on coordinator/pulse calls overlays S1/S2/S3 context (hygiene tags, decision candidates, dimension scores) onto coordinator_lead-style findings; without it you only get the bare finding.
• unacknowledged_only=True filters out findings that have already been triaged via watercooler_daemon_findings(action="acknowledge"), so the skill doesn't repeatedly re-surface the same items run after run.

How to use the findings:

• Scoring input (Step 6): A coordinator_lead or stance_advisory that recurs across multiple threads or repeats over weeks hints at a durable convention worth promoting to a bullet.
• Re-validation input (Step 7): A fresh finding that contradicts an existing bullet (e.g., stalled_handoff repeatedly firing in a domain whose bullet claims the handoff convention is working) is direct evidence the bullet is no longer holding.

Findings are not citations. Daemon emissions are signal inputs, not evidence-backed Watercooler entries. A bullet promoted on the strength of a finding still requires a Decision-entry ULID for its citation slot per Step 8 acceptance rule #4 — trace the finding back to the underlying Decision (often linked via details.source_lead_ids on enriched findings) and cite that.

Step 6: Score and present candidates

For each candidate convention, assign a relevance score:

• High-onboarding: Sourced from a canonical onboarding seed entry's "Active architectural constraints", "Drift findings", or a qualifying Inferred: item (per Step 4.5's four-clause filter — high confidence, code-grounded basis, behavioral-norm framing, durable + cross-cutting). Promote above same-content candidates from other sources — the seed was curated specifically to expose durable architectural truths and carries file:line provenance.
• High: Keywords like must, never, always, convention, rule, naming, workflow in the entry title or body. Appears in a Decision entry.
• Medium: Describes a recurring pattern. Mentioned in multiple entries.
• Low: One-off implementation detail. Unresolved debate. Daemon finding without human follow-up.

Present a scored list with source entry IDs and titles. Ask the user which to incorporate. Do not auto-promote Low candidates.

Risk signals to forward to next risk-register refresh — present this as a separate, parallel section in the candidate report (not mixed in with convention candidates). Sourced from the Step 4.5 (c) scan of seed Inferred: items that qualified as monitoring guidance or recurring-pattern warnings. For each, include:

• Source seed topic + entry_id
• The verbatim Inferred quote
• Confidence level (medium / high)
• Provenance basis (the file, counter, or surface cited)
• Suggested monitoring shape (what to track) and a candidate mitigation (if obvious from the seed body)

These items are informational only — they never auto-patch CLAUDE.md. The user's options are:

1. Forward to the next /watercooler-onboarding refresh to fold into onboarding-risk-register's next entry;
2. Open a dedicated thread to track the risk;
3. Skip — judgment that the signal isn't worth tracking.

If no qualifying risk signals were found, omit this section entirely rather than printing "(none)" — silence here is the common case and shouldn't add noise to the report.

Step 7: Re-validate existing convention bullets

Before adding new bullets, re-read every bullet currently inside the ## Project Conventions block and check that it still describes reality. The generated block is a snapshot of past decisions; underlying code, flags, or APIs may have changed since extraction. Stale bullets actively mislead agents and must be revised or removed in the same patch.

For each existing bullet, verify:

• Named symbols / files / flags exist and behave as described. Grep the source. If a bullet says "registration uses LOCAL_DAEMON_NAMES" and the symbol is gone or marked deprecated, the bullet is stale.
• Source entry has not been superseded. Look for a newer Decision entry that overturns the cited entry. Use either watercooler_list_decisions(topic=<topic>, since=<entry_timestamp>, code_path=<repo_root>) (cleanest — it natively filters by topic and time) or watercooler_search(mode="entries", entry_type="Decision", start_time=<entry_iso>, backend="baseline", code_path=<repo_root>, ...) and inspect the returned entries' timestamp fields manually. Notes: the parameter is entry_type, not type; watercooler_search exposes no sort argument; entry-type filtering is unsupported on the graphiti backend so pass backend="baseline" when filtering by entry_type.
• Referenced docs still say the same thing. If a bullet points to a doc path, open it and confirm the convention still matches.

Categorize each existing bullet as:

State	Action
Still accurate	Keep as-is
Wording stale, intent intact	Revise the bullet, keep the entry ID
Superseded by newer decision	Replace with the new convention; cite the
newer entry ID. Delete the old bullet.
Convention no longer holds	Remove the bullet entirely

Common supersession signals to look for in source code: deprecated, superseded, replaced by, no longer used, # DEPRECATED, an empty return where the symbol used to do work. When any of these is found near a named symbol from a bullet, the bullet is suspect — verify with a fresh search before keeping it.

Present the re-validation result alongside the new candidates from Step 6 so the user sees both adds and revisions/removals in one diff.

Step 8: Patch the generated block

Edit only the ## Project Conventions section of CLAUDE.md. Do not touch any other section.

Hand-added bullets survive re-runs. Any bullet already in the block that carries a valid trailing ULID comment (<!-- 01KXXXXX... -->) was hand-authored against a Decision entry and is subject to Step 7 re-validation like any other bullet — it is not dropped merely because it was hand-added rather than machine-extracted. Only bullets without a ULID (unprovenanced additions) must be removed.

Cap: 20 bullets maximum, ≤ 500 tokens. Each bullet must meet all seven candidate acceptance rules:

1. Durable — likely to hold across multiple PRs or releases

2. Actionable — changes what an agent should do, not just what happened once

3. Cross-cutting — relevant to more than one narrow file or incident

4. Evidence-backed — supported by EITHER (a) a Decision entry, (b) ≥ 2 independent entries, OR (c) a single canonical onboarding-seed entry from architecture / risk-register / docs-contracts / team-map / entry-path (the rule-establishing subset; see Step 4.5). Onboarding seeds carry curated, file:line-cited provenance and were authored specifically as durable foundation material, so they satisfy the auditability bar as a single source. Bullets cited from a seed should typically paraphrase an item from that seed's Observed: (active constraints) or Drift findings: section.

   Include the source identifier in an HTML comment next to the bullet. The identifier must be a Watercooler entry ID matching the regex 01[A-Z0-9]{24} (a ULID like 01KQCJRTFGJPJ75APFG5ESGRJA). PR numbers (PR #685), issue numbers (#694), plan paths (dev_docs/plans/...), and commit SHAs are not entry IDs and must be rejected — find the underlying Decision or seed entry via watercooler_search and cite that ID instead. If no Watercooler entry exists, the convention is not yet evidence-backed and does not belong in this block.

5. Non-duplicative — not already covered by a shorter existing instruction

6. Tier-complete + axis-orthogonal — when the convention names a specific producer, owner, file, flag, or behavior, stress-test by grepping the code for alternate / fallback paths at the same boundary. If an alternate path exists (e.g., a mutex'd open-core fallback for a premium daemon, an else branch handling a different tier, a deprecated symbol replaced by a new one), the bullet must either name both paths or be reframed at a higher abstraction level. Cite a second ULID covering the mutex / fallback / alternate-path contract when both paths exist. Single-path bullets that mask alternate-path reality are not durable — they mislead agents working on the other path.

   Multi-concept claims need a stronger check. When a single bullet equates two or more domain concepts in one statement (e.g. "X uses Y via Z", "open-core ⇔ local backend"), grep each concept independently and verify the equivalence is enforced in code. Many natural-language framings collapse independent axes (transport ↔ backend ↔ distribution; daemon ↔ tier; permission ↔ role) into a single phrase, producing false 1:1 equivalences. If the named axes are independently configurable in the schema or routed by different switches, the equivalence is false — disentangle the axes into separate bullets, or skip and point agents at the authoritative source (config schema, capability table) instead of restating a conflated rule. This applies equally to seed-derived candidates: the seed body is a summary, not a contract — verify the underlying code matches the seed's framing before citing.

7. Provenance-preferred — when both a hand-authored Decision and an ExtractDecisionsDaemon-extracted scribe twin exist for the same content, cite the hand-authored entry. Its body carries the original rationale and author voice; the extracted twin is auto-generated with weaker provenance. Find the source by reading the extracted entry's body — it includes a Source entry: #N <ULID> line naming the original. When in doubt, prefer the entry whose agent field is a human (e.g. Claude Code (jay)) over ExtractDecisionsDaemon (system).

Tier-complete worked example (2026-05-04): A first-pass bullet "stance emission lives on ProjectCoordinatorDaemon, not PulseSnapshotDaemon" failed Rule 6. Grepping src/watercooler_mcp/daemons/__init__.py revealed DecisionStanceDaemon registers as the open-core fallback under a registration-time mutex (lines 631-648). The corrected bullet named both producers and cited a second ULID for the mutex contract. When a single ULID cannot capture both paths, you need two. The first-pass citation also failed Rule 7 — it pointed at the ExtractDecisionsDaemon-extracted scribe twin (01KPV2X1PPVJBK4RST8KTQPFD9) instead of the hand-authored Decision (01KNWMHCNH11EJXGV8X300XP0B) it mirrored.

Axis-orthogonal worked example (2026-05-04): A seed-derived candidate from architecture (01KQN00MXR5VMPD86GHW8E2QHA) proposed "Hybrid memory transport boundary: open-core uses local FalkorDB; hosted uses remote Graphiti via hybrid HTTP transport." Grepping src/watercooler/config_schema.py:1395-1428 revealed three independent axes, not one: (a) transport — execution-routing mode for the local mcp process, values stdio / http / proxy / hybrid (default stdio); (b) backend / extras — T1 baseline always available, T2 Graphiti+FalkorDB opt-in via the [memory] extra, T3 LeanRAG via [leanrag]; (c) distribution — open-core vs hosted Railway deployment. The bullet collapsed all three into a single false equivalence ("open-core ⇔ local FalkorDB"; "hosted ⇔ hybrid transport") that the code does not enforce — open-core users can configure any transport and are baseline-only by default; hosted runs http transport, not hybrid. The bullet was skipped in favor of pointing at the config schema and docs/MCP-CLIENTS.md as authoritative. The seed entry itself had the framing error — Step 4.5's "snapshot in time" warning applies even to high-trust seeds; verify the underlying code matches the seed's summary before citing.

Step 9: Re-derive AGENTS.md and check budget

Repeat Phase 1 Steps 7–8: re-derive AGENTS.md from the updated CLAUDE.md, run wc -l -w, trim if over budget, show the diff, and stop for review.

Update the footer comment to today's date:

<!-- agent-context-refresh:last-updated=YYYY-MM-DD -->

---

Phase 3 — Skill-Surface Audit

Read-only. Detects drift between local skill files and the live MCP tool surface, classifying every reference as OK, DRIFTED, COSMETIC, or NEEDS-REVIEW. Produces a punch list; does not edit any file.

Phase 3 exists because watercooler-cloud tool consolidations (e.g. the 2026-05 PR1a–PR6 arc that took the surface from 52 to 34) retire tool names behind non-discoverable alias forwarders. Per-PR doc sweeps catch most references inside watercooler-cloud, but skill copies in ~/.claude/skills/, downstream repos, and personal CLAUDE.md files routinely drift — silently when an alias still forwards, loudly when it is removed in the next minor release. Phase 3 catches both classes.

Phase 3 reads only public-shippable sources: TOOL_MATRIX (live surface) and TOOL_ALIASES (renamed-with-forwarder map) from src/watercooler_mcp/. Both ship with every open-core install via Copybara's src/** glob, so the audit works identically in private and public contexts. Behavioral-change signals (same name, shifted semantics — e.g. PR2's keyword-search AND→OR default) are intentionally not tracked here; they are handled at retirement time by sweeping affected skill text in the same PR (see PR #828 as the pattern), not by a post-hoc audit.

Step 1: Enumerate the live MCP tool surface

The authoritative list of currently registered tools is whatever the agent's session has loaded. Inside Claude Code, list every tool whose name matches mcp__[a-z0-9_-]+__watercooler_[a-z_]+ from the deferred-tool roster (the "available tools" surface, not just the tools already invoked). Record this as the live set.

If running outside an MCP-aware session, derive the live set from TOOL_MATRIX keys specifically — that dict is the registration-only source of truth for what FastMCP exposes. Do not free-text grep capabilities.py: the file legitimately contains retired tool names inside helper docstrings (e.g. resolve_find_similar_capability() mentions watercooler_find_similar in its docstring even though the name is no longer discoverable), and a bare grep would readmit those names into the live set and silently misclassify stale references as OK.

Preferred — import the dict directly when the package is installed:

python3 -c "from watercooler_mcp.capabilities import TOOL_MATRIX; print('\n'.join(sorted(TOOL_MATRIX.keys())))"

Fallback — AST-parse the file when the package is not importable (e.g. auditing a checkout without an env):

python3 - <<'PY'
import ast, pathlib
tree = ast.parse(pathlib.Path("src/watercooler_mcp/capabilities.py").read_text())
for node in ast.walk(tree):
    if (
        isinstance(node, ast.AnnAssign)
        and isinstance(node.target, ast.Name)
        and node.target.id == "TOOL_MATRIX"
    ):
        for key in node.value.keys:
            if isinstance(key, ast.Constant):
                print(key.value)
        break
PY

Both forms return only the dict keys, so retired names cited in helper docstrings or in _ARG_RESOLVERS keyed by old name cannot leak into the live set. (Aliases are also intentionally excluded — they are forwarders registered by AliasForwardingMiddleware, not entries in TOOL_MATRIX.)

Cross-check: the count should match the surface count reported in the most recent consolidation thread entry (34 after PR6 on 2026-05-22). A mismatch means a registration changed since you last pulled — refresh the checkout before continuing.

Step 2: Load the alias registry

Read TOOL_ALIASES from src/watercooler_mcp/aliases.py. Each key is a retired tool name; each value is a ToolAlias dataclass naming the canonical replacement plus rename_args / inject_args for argument adaptation and the since PR tag. Preferred — import directly when the package is installed:

python3 -c "from watercooler_mcp.aliases import TOOL_ALIASES; \
  import json; print(json.dumps({k: vars(v) for k, v in TOOL_ALIASES.items()}, default=str))"

Fallback — AST-parse the file when the package is not importable (same pattern as Step 1's TOOL_MATRIX fallback). The registry is declared as an annotated assignment (TOOL_ALIASES: dict[str, ToolAlias] = {...}) whose value is a dict literal of <retired_name>: ToolAlias(...) pairs; walk for that shape and extract the four fields Phase 3 uses (canonical, rename_args, inject_args, since):

python3 - <<'PY'
import ast, json, pathlib
tree = ast.parse(pathlib.Path("src/watercooler_mcp/aliases.py").read_text())

def lit(node):
    """Render an ast.Constant / ast.Dict as a JSON-safe value;
    fall back to source for callables (e.g. `guard=lambda args: ...`)."""
    if isinstance(node, ast.Constant):
        return node.value
    if isinstance(node, ast.Dict):
        return {lit(k): lit(v) for k, v in zip(node.keys, node.values)}
    return ast.unparse(node)

aliases = {}
for node in ast.walk(tree):
    if (
        isinstance(node, ast.AnnAssign)
        and isinstance(node.target, ast.Name)
        and node.target.id == "TOOL_ALIASES"
        and isinstance(node.value, ast.Dict)
    ):
        for key_node, val_node in zip(node.value.keys, node.value.values):
            if not (isinstance(key_node, ast.Constant)
                    and isinstance(val_node, ast.Call)):
                continue
            aliases[key_node.value] = {
                kw.arg: lit(kw.value)
                for kw in val_node.keywords
                if kw.arg in {"canonical", "rename_args", "inject_args", "since"}
            }
        break

print(json.dumps(aliases, indent=2, default=str))
PY

TOOL_ALIASES is the source of truth for live alias forwarding (it's what AliasForwardingMiddleware reads at request time). It covers the only class of drift Phase 3 needs to label specifically: names that still call through but are scheduled for removal one minor release later.

Retirements with no forwarder ("the name is just gone") are deliberately not enumerated anywhere — when one of those names appears in a skill file, Step 4 surfaces it as NEEDS-REVIEW and the user investigates by hand (read aliases.py git history, or ask a maintainer).

Step 3: Scan skill files for tool references

Grep these locations (paths exist when applicable):

• ~/.claude/skills/**/SKILL.md and **/SKILL.public.md — user-level skills
• ~/.claude/CLAUDE.md — user-global instructions
• <repo_root>/.claude/skills/**/SKILL.md — project-level skills
• <repo_root>/CLAUDE.md, <repo_root>/AGENTS.md — project agent context

Pattern (single grep, both forms):

grep -rnE "(mcp__[a-z0-9_-]+__)?watercooler_[a-z_]+" \
  ~/.claude/skills/ ~/.claude/CLAUDE.md \
  .claude/skills/ CLAUDE.md AGENTS.md 2>/dev/null

Strip the mcp__<server>__ prefix so each hit is a bare watercooler_* name comparable to the live set and manifest keys.

Step 4: Classify each hit

For every grep result, look up the bare name in (live set, TOOL_ALIASES) and apply this classification:

Classification	Meaning	Trigger
OK	Name is in the live set	No action needed — skip from the report
DRIFTED	Call works via alias forwarder; will break when the alias is removed (one minor release later)	Name absent from live set AND present in TOOL_ALIASES. Report the canonical replacement plus any rename_args / inject_args from the alias entry.
COSMETIC	Historical reference in audit/inventory/change-log text — not a call site	Name present in TOOL_ALIASES AND the surrounding line is explanatory (e.g. inside a table comparing old→new, inside .claude/skills/watercooler-tool-audit/, or framed as past tense — "was replaced by", "is now"). Use judgment; when ambiguous, prefer DRIFTED.
NEEDS-REVIEW	Name absent from both the live set and TOOL_ALIASES	Either a retirement that landed without a forwarder, a typo, or a name from an unfamiliar fork. The user investigates by hand — read src/watercooler_mcp/aliases.py git history, search the consolidation thread, or ask a maintainer. The audit cannot disambiguate further from public sources alone.

Behavioral-change tracking is intentionally not part of Phase 3. Shifts that retain a tool's name (e.g. PR2's AND→OR keyword-search default, PR1b's authority resolution) are addressed at retirement time by sweeping affected skill text in the same PR — see PR #828 as the established pattern.

Step 5: Report

Render the result as a single markdown table sorted by file:line, plus a per-file rollup with DRIFTED / COSMETIC / NEEDS-REVIEW counts and a one-line suggested fix per file. For DRIFTED hits, include the canonical replacement and any rename_args / inject_args the caller will need.

Table columns: File, Line, Hit, Class, Canonical, Notes (the alias entry's since PR tag, when present).

Stop here. Phase 3 does not edit any file. The punch list is the deliverable; the user decides whether to patch in-place or wait for the next downstream skill-sync.

Step 6: Footer (rarely needed)

If Phase 3 was the sole reason for the run and it surfaced no actionable drift in CLAUDE.md / AGENTS.md, leave the agent-context-refresh last-updated footer unchanged. CLAUDE.md/AGENTS.md staleness is the domain of Phase 1's drift-fact step; Phase 3's domain is the wider skill ecosystem.

---

Degraded Paths

Condition	Phase 1	Phase 2	Phase 3
Watercooler unavailable	Proceed with static refactor; skip role fetch; note limitation	Stop with clear error — no evidence, no extraction	No impact — Phase 3 reads source files only, not the graph
T2 unavailable (open-core)	No impact	Continue with T1 evidence only; skip smart_query	No impact
No decisions since last update	No impact	Report "no new candidates found"; skip patch	n/a
No onboarding seeds present (no canonical topics from watercooler-onboarding exist)	Skip the seed reads in Step 4; note in the run report that running /watercooler-onboarding would strengthen the next refresh	Skip Step 4.5; report the gap in the candidate list and recommend running /watercooler-onboarding before the next Phase 2	No impact
From-scratch seed mode + no onboarding seeds (hardest case — no CLAUDE.md, no seeds)	Build Repository Conventions / Development Workflow / Security & Publishing from source files and observed codebase patterns; mark every unsourced micro-convention inferred in the Step 8 diff; recommend /watercooler-onboarding then a Phase 2 run to firm them up	n/a	No impact
Outside Claude Code (no deferred-tool roster to enumerate the live set)	n/a	n/a	Fall back to the TOOL_MATRIX and TOOL_ALIASES AST/import paths in Steps 1 and 2; note the fallback in the report header

---

Conventions Not to Touch

Phase 2 should reject or leave in Watercooler when the candidate is:

• A one-off implementation detail
• An unresolved debate
• A daemon finding without human follow-up
• A policy that conflicts with current repo docs or code
• Already covered by a shorter existing instruction