AI Agents Docs Of Record

ai-agents-docs-of-record

This repo runs on verification-based governance. Continuity lives in per-issue handoffs and Serena memory. Session log creation is discontinued; transcript and pull request evidence remain valid, and the ~1,500 historical logs under .agents/sessions/ stay in the repo as history.

Triggers

• write the session log
• create an ADR
• fill the retro
• save this to memory
• which document of record does this go in

The Record System Map

Process

Phase 1: Pick the right record

One fact can need two records: an incident gets a retrospective AND a memory distilling the rule. The retro is the evidence; the memory is the searchable lesson.

Phase 2: Session logs (historical only)

Session log creation is discontinued (.claude/rules/session-logs.md MUST 1): do not create a new .agents/sessions/*.json file. The ~1,500 existing logs stay in the repo as history, schema at .agents/schemas/session-log.schema.json (top-level required: session, protocolCompliance with sessionStart and sessionEnd subsections), and remain readable by retrospective, memory extraction, and the PreCompact hook.

If a historical log is staged anyway (for example, cherry-picked from an older branch), it still needs to pass:

uv run python3 scripts/validate_session_json.py .agents/sessions/<file>.json

Hard requirements enforced by scripts/validate_session_json.py:

• Required session fields: number, date, branch, startingCommit, objective.
• Branch must match ^(feat|fix|docs|chore|refactor|test|ci)/.
• startingCommit must be a 7-40 char lowercase hex SHA.

Historical naming may carry a slug: 2026-06-30-session-2600-skill-tooling-hygiene-taste-lints-naming.json. Never resolve a session-log merge conflict with git checkout --ours; the settled rule from the session 1187 incident is take --theirs and rename your own log to the next free number (see ai-agents-failure-archaeology).

Phase 3: ADRs

Numbering collides here; treat numbers as claims, not identity. History: concurrent ADR PRs each picked the same "next" number (Issue #2253), and ADR-058/062/063 each existed twice until the duplicates were renumbered to 069/070/071 (Issue #2228, per the docstring of scripts/validation/check_adr_uniqueness.py). ADR-055 and ADR-024 share the slug github-actions-runner-selection (055 supersedes 024). Rule: when citing an ADR, verify the file content says what you claim, not just the number.

python3 scripts/validation/check_adr_uniqueness.py --print-next

Workflow:

1. Search first: memory-search for prior decisions, then read neighbors in .agents/architecture/ (ADR-007 memory-first: retrieval precedes reasoning).
2. Generate with adr-generator against .agents/architecture/ADR-TEMPLATE.md.
3. Claim the number from --print-next at the last moment before commit.
4. Expect the gate: any ADR-*.md create or edit fires the adr-review multi-agent debate (AGENTS.md, "ADR Review"). Governance-level changes additionally need human approval (AGENTS.md Boundaries: "Ask First: Architecture|New ADRs|Breaking|Security").
5. Lifecycle frontmatter per ADR-073 (accepted 2026-06-19): optional, unenforced Phase-1 fields id, status, date, decision-makers, supersedes, superseded-by, explainer, implemented. Use them on new ADRs; do not backfill mechanically.

Phase 4: Retrospectives

Two skills, different jobs. reflect captures corrections and praise mid-conversation while context is hot. retrospective produces the full artifact (timeline, Five Whys, learning matrix) in .agents/retrospective/.

Retrospectives are written on demand: /retro here, plus the Post-PR Retrospective workflow in CI. Until #3349 a Stop hook also wrote a skeleton at session end stamped <!-- RETRO-STATE: skeleton-pending-fill --> (Issue #2079); that hook is deleted, but skeletons it already wrote are still placeholders rather than records, so fill them with /retro fill YYYY-MM-DD.

INDEX.md was auto-appended by that hook (Issue #1703) and is incomplete: 5 rows against 95 retro files as of 2026-07-03. Nothing appends to it now. Never treat INDEX.md as the catalog; list the directory. Also note retro-cited short SHAs do not resolve locally even with full history present (~1471 commits as of 2026-07-03), so retros and memories, not git archaeology, are the durable history (depth in ai-agents-failure-archaeology).

Retro learnings MUST be persisted to Serena in the same session, not left as prose in the artifact (.serena/memories/retrospective-accuracy.md: "learnings were proposed in the retrospective but then not persisted to Serena, so they will be lost").

Phase 5: Serena memories

Write conventions (canonical text: .claude/skills/memory/SKILL.md, "Serena Write Conventions"):

• File name: [domain]-[descriptive-name].md, lowercase, hyphens. Example: pr-review-security.md.
• Entity ID inside the file: {domain}-{description}, kebab-case, no prefix. File name and entity ID are separate things; do not conflate them.
• Domain index files (skills-*-index.md) are a two-column table ONLY. Insert new rows AFTER the last existing data row, never after the header or delimiter row.
• Register new memories where memory-index (the keyword-to-memory routing map read at session start) will find them.

Curation (marking obsolete, dedupe, bidirectional links) is the curating-memories skill. Search is memory-search (Tier 1); the four-tier model is ADR-063.

Separate-PR rule: .claude/rules/claude-agents.md MUST NOT item 2: "MUST NOT bundle skill code changes with memory changes in the same PR (separate concerns)." Memory edits ride their own PR. The cautionary tale is PR #908, where 53 memory files changed in a PR that started as one CI fix (.agents/retrospective/2026-01-15-pr-908-comprehensive-retrospective.md).

Phase 6: Handoffs

Tier model per ADR-014 and .agents/sessions/handoffs/README.md:

Start from the template .agents/templates/HANDOFF.md. A handoff must carry a "Verification on Resume" checklist; the next session executes it before any Edit or Write (templates/agents/implementer.shared.md, "Session Start", a blocking gate). The old project-wide .agents/HANDOFF.md dashboard is deleted: it had grown to 122KB with an 80%+ per-PR merge conflict rate before ADR-014 froze it, and per-issue handoffs plus Serena memory now carry the continuity it used to (.agents/architecture/ADR-014-distributed-handoff-architecture.md:13-16).

Phase 7: House style and stale docs

Prose rules (canonical: .claude/rules/voice.md, .claude/rules/universal.md):

• Lead with the point. Name file, line, command, number, or say you do not know.
• No em-dashes or en-dashes anywhere in authored text (universal.md MUST NOT item 4; only tests/hooks/fixtures/ is exempt). Every dash costs a bot review thread.
• No banned vocabulary (voice.md "Banned Vocabulary" section).
• No auto-generated headers or timestamps-as-headers in any authored or generated file (universal.md MUST NOT item 5; rejected by the user three times as of 2025-12-17).
• Block-style YAML arrays only in frontmatter (.agents/governance/PROJECT-CONSTRAINTS.md, Frontmatter constraints table; ADR-040 Amendment). Inline ['a', 'b'] breaks Copilot CLI parsing.
• Run prose-self-check before emitting any session-log narrative, ADR context section, retro, or PR description.

Markdown lint is SCOPED to files you changed. The PR #908 Five Whys traced 53 unrelated memory-file changes to markdownlint --fix **/*.md run repo-wide (retro, Q1-Q4). Lefthook filters staged Markdown and runs the named validator declared in lefthook.yml. Config is .markdownlint-cli2.yaml. Never widen the glob.

Mirror claims (FM-9): any statement that a file "matches", "mirrors", or "aligns with" another source MUST cite the canonical path and quote the load-bearing fragment character-for-character; intentional divergence gets a named divergence section (.agents/governance/FAILURE-MODES.md, failure mode 9; rule text in .claude/rules/canonical-source-mirror.md; heuristic gate scripts/validation/check_canonical_citations.py, warn-only unless STRICT_CANONICAL_CHECK=1). Cost of ignoring it: PR #1887 took 7 fix commits on confident-incorrectness claims.

Fix stale docs on contact. Worked example from before PR #2871: CONTRIBUTING.md line 155 said build/Generate-Agents.ps1 PowerShell invocation until PR #2871 repointed it to build/generate_agents.py. Zero .ps1 build scripts exist (ADR-042 Python migration). When you touch a stale doc: fix the lines on your path, quote the canonical source per FM-9, and flag (do not silently expand into) the rest. Behavioral claims in docs are verified with doc-accuracy.

Anti-Patterns

Verification

Before ending any session that produced records:

• Any staged or supplied session log passes uv run python scripts/validate_session_json.py <log>.
• Any new ADR number came from check_adr_uniqueness.py --print-next and the adr-review gate ran.
• Retro learnings are persisted as Serena memories, not just prose in the retro file.
• New memories follow [domain]-[name].md naming and their index row sits after the last data row.
• Open-issue work has a per-issue handoff with a Verification on Resume checklist.
• grep -rPn '[\x{2013}\x{2014}]' <changed .md files> returns nothing; lint ran scoped to changed files only.
• Every "matches/mirrors" claim quotes its canonical source verbatim.

Provenance and Maintenance

Verified 2026-07-03 against the working tree. Re-verify volatile facts before relying on them:

Maintenance: when a validator, template path, or protocol phase changes, update the matching row here in the same PR. This file is itself a document of record; it obeys everything it prescribes.