Spec graph

spec-graph: Agent Operator Skill

spec-graph is a CLI tool that layers a typed semantic graph on top of phase-based development. The structured graph — not markdown — is the source of truth. Agents receive computed impact sets and patch-target lists instead of reasoning over free-text documents.

Four core capabilities:

• Impact Analysis — compute what must change together when an entity changes
• Gap Detection — find missing implementations, verifications, plans, or open questions
• Consistency Validation — check graph integrity and workflow gates
• Agent Coordination — work only on computed affected targets, not entire documents

Three-Layer Architecture

v1 organizes the graph into three distinct layers. Each layer has its own entity types, relation types, edge matrix, and validation checks.

arch (architecture layer)

Contains the "what" and "why" of the system: requirements, decisions, interfaces, states, tests, and supporting entities. This is where semantic meaning lives.

Entity types: requirement, decision, interface, state, test, crosscut, criterion, assumption, risk, question

Relation types: implements, verifies, depends_on, constrained_by, triggers, answers, assumes, has_criterion, mitigates, supersedes, conflicts_with, references

exec (execution layer)

Contains the "when" and "how" of delivery: plans, phases, tasks, and changes. A plan groups phases into a single active delivery sequence. Only one plan may be active at a time. A change is a lightweight independent work unit (PR, bugfix, patch) that covers arch entities without belonging to any plan or phase.

Entity types: plan, phase, task, change

Relation types: belongs_to exactly for phase→plan and task→phase, task_depends_on (dependent task→prerequisite task in the same phase), precedes (phase→phase), and blocks (phase→phase).

Note: change entities do NOT participate in exec relations (belongs_to, precedes, blocks). They are independent units.

mapping (cross-layer)

Connects arch entities to exec entities. This is where intent meets delivery.

Relation types: covers (phase/change/task→arch entity), delivers (phase/task→arch entity). When a phase has any child task, task mappings are canonical and direct phase mappings are forbidden.

Layer Classification

Layer is determined by entity type prefix. It is always deterministic:

Prefix	Type	Layer
REQ	requirement	arch
DEC	decision	arch
API	interface	arch
STT	state	arch
TST	test	arch
XCT	crosscut	arch
ACT	criterion	arch
ASM	assumption	arch
RSK	risk	arch
QST	question	arch
PLN	plan	exec
PHS	phase	exec
TSK	task	exec
CHG	change	exec

---

Core Principles

0. English only: all spec-graph content — entity titles, descriptions, metadata, --reason messages, criteria text, and any other field written into the graph — MUST be in English. Regardless of the language used in conversation, never write non-English text into spec-graph entities or relations.
1. Compute first: never modify by guesswork. Always run impact and validate to identify targets before making changes.
2. JSON contract: all CLI output goes to JSON stdout. Parse it to decide the next action.
3. Layer discipline: arch entities belong in arch, exec entities in exec. Do not mix concerns.
4. Phase gates: always run validate --layer mapping --phase before starting or completing a phase.
5. Git as audit log: each commit is a logical changeset. The project's git history is the sole audit trail for spec-graph changes.
6. covers/delivers: use the v1 mapping relations. covers expresses planning intent, delivers expresses completion.
7. Graph-native plans: new plans create TSK entities and no Markdown plan files. Never auto-import, delete, or reinterpret old Markdown.

Task Contract and Scope

Every task has a non-empty title and description plus a closed six-field metadata contract. Unknown keys are rejected:

{
  "order": 1,
  "instructions": ["Implement the scoped behavior."],
  "acceptance": ["The behavior is verified."],
  "must_not": [],
  "references": [],
  "qa": [{"command":"go test ./...","expected":"exit 0","evidence":""}]
}

• order is a positive integer.
• instructions, acceptance, and qa are non-empty arrays.
• must_not and references are required arrays and may be empty.
• Every QA item has non-empty command and expected. evidence is empty before resolution and must identify a repository-relative regular file when the task resolves.

Canonical scope is selected per phase: a task-managed phase (any task belongs_to it) derives scope and delivery from the union of child task covers/delivers; a taskless phase keeps its direct phase mappings unchanged. Never mix direct phase mappings with child-task mappings. Tasks remain exec entities and are never members of the architecture satisfaction closure.

Task lifecycle is draft → active → resolved or draft|active → deprecated; resolved and deprecated are terminal, and deprecation requires a reason. Activation requires an active parent phase and resolved prerequisites. Resolution requires resolved prerequisites, QA evidence, and delivers for every deliverable target the task covers.

---

Storage Architecture

v0.3.0 uses TOML-first storage with SQLite as a disposable index.

• Source of truth: TOML files at .spec-graph/entities/{type}/{id}.toml
• Relations: embedded in entity TOML files (outbound only)
• SQLite index: .spec-graph/graph.db, disposable, auto-rebuilt from TOML on any command if stale
• Staleness detection: content-hash fingerprint per entity file
• Gitignored: .spec-graph/graph.db* and .lock are never committed

The SQLite index exists purely for fast queries (neighbors, impact, path). If deleted or corrupted, it rebuilds automatically. Never treat it as authoritative.

TOML File Format

Canonical entity file at .spec-graph/entities/requirement/REQ-001.toml:

schema = 1
id = "REQ-001"
type = "requirement"
title = "User authentication"
description = "All APIs require OAuth2"
status = "active"
created_at = 2026-05-23T17:00:00+09:00
updated_at = 2026-05-23T17:30:00+09:00

[metadata]
priority = "must"
kind = "non_functional"

[[relations]]
to = "ACT-001"
type = "has_criterion"

[[relations]]
to = "DEC-001"
type = "constrained_by"
weight = 0.8

Fields: schema (always 1), id, type, title, description (optional), status, created_at, updated_at, [metadata] (type-specific), [[relations]] (outbound edges).

Git Workflow

TOML files are designed for git-friendly collaboration:

• Each entity is a separate file, so merge conflicts are entity-scoped
• After git merge or git pull with conflicts, resolve TOML files then run spec-graph doctor
• SQLite index is never committed (listed in .gitignore)
• Commit messages serve as the audit log. Git history is the sole mechanism for tracking entity changes over time.

Audit Trail

spec-graph does NOT maintain its own history. The project's git history is the sole audit trail.

• Entity changes: git log -- .spec-graph/entities/{type}/{id}.toml
• Relation changes: tracked via the owning entity's file history
• Phase transitions: git log -- .spec-graph/entities/phase/PHS-XXX.toml

Recommendation: Commit .spec-graph/ changes after each logical unit of work (phase activation, delivers batch, entity registration).

---

Quick Reference: CLI Commands

See references/cli-reference.md for full options.

Project Init

spec-graph init
spec-graph init --path /custom/path

Entity CRUD

spec-graph entity add --type <TYPE> --title "..." [--id <ID>] [--description "..."] [--metadata '{}']
# --id is optional. When omitted, auto-generated (MAX+1 for that type). Capture returned .entity.id via jq -r '.entity.id'.
spec-graph entity get <ID>
spec-graph entity list --type <TYPE> [--status <STATUS>] [--layer arch|exec|mapping|all]
spec-graph entity update <ID> --title "..."
spec-graph entity update <ID> --status resolved [--force]
spec-graph entity deprecate <ID>
spec-graph entity delete <ID>

Relation CRUD

spec-graph relation add --from <ID> --to <ID> --type <RELATION_TYPE>
spec-graph relation list --from <ID> [--layer arch|exec|mapping|all]
spec-graph relation delete --from <ID> --to <ID> --type <RELATION_TYPE>

Impact Analysis

spec-graph impact <ID> [<ID>...]
spec-graph impact <ID> --follow implements,verifies,covers
spec-graph impact <ID> --min-severity medium
spec-graph impact <ID> --dimension structural|behavioral|planning
spec-graph impact <ID> --layer arch

Validation

spec-graph validate
spec-graph validate --layer arch
spec-graph validate --layer exec
spec-graph validate --layer mapping
spec-graph validate --check orphans|coverage|cycles|conflicts|invalid_edges|superseded_refs|unresolved
spec-graph validate --check phase_order|single_active_plan|orphan_phases|exec_cycles|invalid_exec_edges
spec-graph validate --check plan_coverage|delivery_completeness|mapping_consistency|invalid_mapping_edges
spec-graph validate --phase <PHS-ID>
spec-graph validate --entity <ID>

Query

spec-graph query neighbors <ID> --depth 2
spec-graph query path <FROM-ID> <TO-ID>
spec-graph query scope <PHS-ID>
spec-graph query unresolved --type question|assumption|risk [--phase <PHS-ID>]
spec-graph query sql "SELECT ..."

Phase Lifecycle

spec-graph phase next [--activate]
spec-graph phase context <PHS-ID>

Export

spec-graph export --format json|dot|mermaid
spec-graph export --center <ID> --depth 3 --format json
spec-graph export --layer arch --format dot

Bootstrap

spec-graph bootstrap scan --input ./docs/ [--format json]
spec-graph bootstrap import --input extracted.json --mode review

Migration & Integrity

# One-shot migration from old SQLite-only format
spec-graph migrate [--dry-run] [--keep-db]

# Integrity validation (run after git merge/pull)
spec-graph doctor [--check <name>] [--fix]

---

Entity & Relation Quick Reference

See references/data-model.md for full type catalog, metadata schemas, and edge matrices.

Entity Types (14)

Prefix	Type	Layer	Purpose
REQ	requirement	arch	functional / non-functional requirement
DEC	decision	arch	policy / architecture decision
API	interface	arch	API contract, module interface, event contract
STT	state	arch	state or state-transition rule
TST	test	arch	test case / scenario
XCT	crosscut	arch	cross-cutting concern (auth, audit, etc.)
QST	question	arch	unresolved question
ASM	assumption	arch	unverified assumption
ACT	criterion	arch	acceptance criterion
RSK	risk	arch	explicit risk item
PLN	plan	exec	delivery plan grouping phases
PHS	phase	exec	development phase or milestone
TSK	task	exec	graph-native unit of implementation work
CHG	change	exec	lightweight work unit (PR, bugfix, patch)

Entity Status: draft → active → deprecated / resolved / deleted

Auto-activation (v0.4.0+): When a delivers relation is added targeting an arch entity, the CLI automatically transitions that entity from draft to active. This means:

• draft = registered, no delivery evidence yet
• active = at least one phase has delivered this entity (auto or manual)
• resolved = verified complete (only spec-verifier should set this)

Do NOT manually transition arch entities to active after adding delivers — the CLI handles it. Do NOT expect delivers to auto-resolve entities; resolution requires explicit verification.

Gated transitions (v0.3.1+): Transitioning a phase or plan to resolved is gated. The CLI automatically runs delivery_completeness + gates checks (for phases) or plan_coverage (for plans). If issues are found, the transition is blocked (exit 2). Completion findings can be accepted with --force --reason "..."; structural findings cannot be bypassed. A successful forced completion emits outcome = "applied_with_force", reports the accepted findings, and records top-level completion_forced = true plus completion_reason = "..." in the entity TOML. A blocked transition emits outcome = "blocked" and exits 2. Phase gate checks evaluate only the target phase, its child tasks, and their effective mapping scope; unscoped validate remains graph-wide.

PLN / PHS Lifecycle

Status State Machine

PLN:  draft → active → resolved (gated: plan_coverage)
                     → deprecated (--force required)

PHS:  draft → active → resolved (gated: delivery_completeness + gates)
                     → deprecated (--force required)

Transition Ownership

Transition	Owner	Precondition
PLN: draft → active	spec-planner	Only one active plan allowed
PHS: draft → active	spec-executor	Predecessor phases resolved (soft — warn if not)
PHS: active → resolved	spec-verifier	All deliverables verified, gate passes
Any → deprecated	User (manual)	--force required

Rules

1. Only one active PLN at a time — single_active_plan check enforces this.
2. PHS activation order: phases with precedes predecessors should be activated in order. Activating out-of-order is allowed but triggers a warning.
3. PHS resolution is gated: entity update PHS-XXX --status resolved auto-runs delivery_completeness + gates. Blocked (exit 2) if issues exist.
4. PLN resolution is gated: requires plan_coverage — all active arch entities must be covered.
5. No skipping states: draft → resolved is invalid. Must pass through active.
6. deprecated is terminal: no transitions out of deprecated.

Relation Types (18)

Architecture layer (12): implements, verifies, depends_on, constrained_by, triggers, answers, assumes, has_criterion, mitigates, supersedes, conflicts_with, references

Execution layer (4): belongs_to, task_depends_on, precedes, blocks

Mapping layer (2): covers, delivers

---

Agent Workflow Patterns

This section is the heart of this skill. Agents follow these patterns.

Pattern 1: Plan and Phase Setup

Create a graph-native plan, add phases and tasks, then wire task scope. This path creates no Markdown. Direct phase mappings shown in older projects are a legacy taskless path only.

Note: This pattern uses explicit --id because it's a batch/cross-referencing flow where IDs are referenced before all entities exist. For single interactive creates, you can omit --id and capture the auto-generated ID from .entity.id (via jq -r '.entity.id').

# 1. Create the plan (only one active plan allowed)
spec-graph entity add --type plan --id PLN-001 \
  --title "v1 Delivery Plan" \
  --metadata '{"status":"active"}'

# 2. Create a phase
spec-graph entity add --type phase --id PHS-001 \
  --title "Phase 1 - Auth" \
  --metadata '{"goal":"Build authentication","order":1,"exit_criteria":["Auth API complete","E2E tests pass"]}'

# 3. Assign the phase to the plan
spec-graph relation add --from PHS-001 --to PLN-001 --type belongs_to

# 4. Create tasks with the closed TaskContract
spec-graph entity add --type task --id TSK-001 --title "Implement auth API" \
  --description "Implement the authentication API and tests." \
  --metadata '{"order":1,"instructions":["Implement the auth API."],"acceptance":["Auth tests pass."],"must_not":[],"references":[],"qa":[{"command":"go test ./...","expected":"exit 0","evidence":""}]}'
spec-graph entity add --type task --id TSK-002 --title "Integrate auth flow" \
  --description "Integrate the completed authentication API." \
  --metadata '{"order":2,"instructions":["Integrate the auth flow."],"acceptance":["Integration tests pass."],"must_not":[],"references":[],"qa":[{"command":"go test ./...","expected":"exit 0","evidence":""}]}'

# 5. Wire membership, dependency, and canonical task scope
spec-graph relation add --from TSK-001 --to PHS-001 --type belongs_to
spec-graph relation add --from TSK-002 --to PHS-001 --type belongs_to
spec-graph relation add --from TSK-002 --to TSK-001 --type task_depends_on
spec-graph relation add --from TSK-001 --to REQ-001 --type covers
spec-graph relation add --from TSK-002 --to REQ-002 --type covers

# 6. Validate and obtain the executor/verifier contract
spec-graph validate --layer exec --check task_graph
spec-graph validate --layer mapping --phase PHS-001 --check task_scope
spec-graph phase context PHS-001

phase context returns {plan,phase,tasks,scope,delivery,blockers,ready_task_ids,blocked_task_ids}. Each task entry is {entity,contract,prerequisite_ids,covers,delivers}. The same result is available as RPC phase.context and MCP phase_context.

Pattern 2: Change Handling

When an existing entity changes, always run impact first:

# 1. Compute impact — what else must change
spec-graph impact DEC-031 --dimension behavioral

# 2. Inspect affected targets (parse JSON)
spec-graph impact DEC-031 | jq '.affected[] | {id, type, impact, reason}'

# 3. Check unresolved items
spec-graph query unresolved --type question

# 4. Modify only affected targets (do not touch unrelated entities)
spec-graph entity update DEC-031 --title "New decision"

# 5. Full validation
spec-graph validate

Never modify related entities by guesswork without running impact first.

Pattern 3: Phase Exit

Phase completion is gated by the CLI (v0.3.1+). Running entity update --status resolved automatically enforces delivery_completeness + gates checks. If issues exist, the transition is blocked with exit code 2.

# Direct completion attempt — gate runs automatically
spec-graph entity update PHS-002 --status resolved

# If blocked, resolve issues first:
# 1. Review graph-native phase context
spec-graph phase context PHS-002

# 2. Check what's missing
spec-graph validate --layer mapping --phase PHS-002 --check delivery_completeness
spec-graph validate --layer mapping --phase PHS-002 --check gates

# 3. Fix issues (add delivers, answer questions, mitigate risks)
spec-graph relation add --from TSK-005 --to REQ-001 --type delivers

# 4. Retry
spec-graph entity update PHS-002 --status resolved

# Force completion findings only (structural findings remain blocked)
spec-graph entity update PHS-002 --status resolved --force \
  --reason "Accept the documented completion risk"

Update responses expose one of three outcomes: applied, applied_with_force, or blocked. blocked always means the TOML was left unchanged; a non-nil gate report does not by itself mean the update was blocked.

Pre-flight checks (optional, for visibility before attempting completion):

# Review scope
spec-graph query scope PHS-002

# Arch coverage
spec-graph validate --layer arch --check coverage

# Mapping consistency
spec-graph validate --layer mapping --phase PHS-002 --check mapping_consistency

# Exec ordering
spec-graph validate --layer exec --check phase_order

If validate reports issues, resolve them before attempting --status resolved.

Handling "covered but not delivered" mapping failures

When delivery_completeness reports a covered arch entity has no delivers relation:

# 1. Identify what the phase covers
spec-graph query scope PHS-002

# 2. Find implementing entities for the requirement
spec-graph relation list --to REQ-001   # find what implements/verifies REQ-001

# 3. Determine the MINIMAL proxy set — only entities whose delivery in this
#    phase is necessary and sufficient to consider REQ-001 delivered
#    Ask: "Which implementing entities are necessary and sufficient?"

# 4. Add delivers ONLY for that minimal set
spec-graph relation add --from PHS-002 --to API-005 --type delivers
spec-graph relation add --from PHS-002 --to TST-001 --type delivers

# 5. Re-validate
spec-graph validate --layer mapping --phase PHS-002 --check delivery_completeness

Critical rules for delivery proxy resolution:

• Compute the minimum set of implementing entities per requirement. Do not add all related entities.
• If the check still fails after adding the minimal proxy set, investigate the validator semantics or the graph model before expanding further. Do not blindly widen the delivered set.
• Apply the same precision level consistently across all phases.
• After adding proxy relations, verify semantic correctness: does each delivers accurately represent work completed in this phase, or is it just silencing the check?

Pattern 4: Full Patch Orchestration (recommended)

The safest change-handling flow:

1. Identify change target
2. spec-graph impact → compute affected set
3. spec-graph validate → check currently broken rules
4. Modify only affected targets (entity update, relation add/delete, etc.)
5. Semantic review → does each added relation accurately represent the intended meaning?
6. spec-graph validate → re-verify after modifications

The agent modifies only entities in the affected list from step 2. If an entity outside the list needs modification, first run query neighbors to verify the relationship.

Step 5 (semantic review) is critical. Before re-validating, review every relation you added and ask: "Does this relation reflect a real semantic relationship, or am I adding it to pass a check?" Check passage alone does not prove graph correctness. A graph that passes all checks but contains over-broad relations is worse than one that fails a check with an honest gap.

Pattern 5: Adding a Requirement

Typical flow for adding a new requirement and wiring it into the graph:

Tip: For a single ad-hoc requirement, you can omit --id and capture the returned ID: REQ_ID=$(spec-graph entity add --type requirement --title "..." --metadata '...' | jq -r '.entity.id'). The example below uses explicit IDs for clarity.

# 1. Create requirement
spec-graph entity add --type requirement --id REQ-015 \
  --title "All payments must be idempotent" \
  --metadata '{"priority":"must","kind":"non_functional","owner":"payment-team"}'

# 2. Attach acceptance criterion
spec-graph entity add --type criterion --id ACT-020 \
  --title "Duplicate request within window processed only once" \
  --metadata '{"given":"Payment request already sent","when":"Same request resent","then":"No duplicate processing; return existing result"}'
spec-graph relation add --from REQ-015 --to ACT-020 --type has_criterion

# 3. Map to phase using covers (not planned_in)
spec-graph relation add --from PHS-003 --to REQ-015 --type covers

# 4. Link crosscut constraint (if applicable)
spec-graph relation add --from REQ-015 --to XCT-002 --type constrained_by

# 5. Validate arch layer
spec-graph validate --layer arch --entity REQ-015

Pattern 6: Bootstrap (graph from existing docs)

When existing markdown documents are available:

# 1. Extract candidates — generates review candidates, not auto-committed
spec-graph bootstrap scan --input ./docs/ --format json

# 2. Review — filter low-confidence items
cat extracted.json | jq '.entities[] | select(.confidence >= 0.7)'

# 3. Import in review mode
spec-graph bootstrap import --input extracted.json --mode review

Low-confidence relations are never auto-imported. A human must confirm, or the agent must cross-reference against the source document before deciding.

---

Validation Checks Guide

When to use each check. See references/validation-rules.md for detailed rules.

Architecture Layer Checks (--layer arch)

Check	Purpose	When to Run
orphans	isolated arch entities with no relations	periodic cleanup, before phase start
coverage	missing implementations / tests	required before phase exit
cycles	circular references in depends_on chains	after adding relations
conflicts	semantic conflicts between entities	after changes
invalid_edges	arch edge matrix violations	after adding relations
superseded_refs	active refs to deprecated entities	after deprecation
unresolved	open questions, unverified assumptions, unmitigated risks	before phase start

Execution Layer Checks (--layer exec)

Check	Purpose	When to Run
phase_order	phases with precedes/blocks form a valid sequence	after adding exec relations
single_active_plan	only one plan is active	after plan creation or status change
orphan_phases	phases not belonging to any plan	after adding phases
exec_cycles	circular precedes/blocks chains	after adding exec relations
invalid_exec_edges	exec edge matrix violations	after adding exec relations
orphan_changes	changes with no relations to other entities	after adding changes
task_graph	task parents, same-phase dependencies, and cycles	after changing tasks or dependencies

Mapping Layer Checks (--layer mapping)

Check	Purpose	When to Run
plan_coverage	all active requirements are covered by some phase	before phase start
delivery_completeness	covered arch entities have delivery evidence	auto-enforced on phase → resolved
mapping_consistency	covers/delivers targets exist and are arch entities	after adding mapping relations
invalid_mapping_edges	mapping edge matrix violations	after adding mapping relations
gates	unresolved questions, unmitigated risks, draft decisions	auto-enforced on phase → resolved
task_scope	task coverage, delivery subset, and no mixed mappings	after changing task mappings

For task-managed completion, four checks/gates matter: structural task_graph, mapping task_scope, task completion delivery/evidence gates, and phase child-resolution plus existing delivery_completeness/gates checks.

Common Combinations

# Before phase start
spec-graph validate --layer exec --check single_active_plan
spec-graph validate --layer exec --check phase_order
spec-graph validate --layer arch --check unresolved
spec-graph validate --layer mapping --check plan_coverage

# Before phase completion (now auto-enforced by entity update --status resolved)
# These are still useful for pre-flight visibility:
spec-graph validate --layer arch --check coverage
spec-graph validate --layer mapping --phase PHS-003 --check delivery_completeness
spec-graph validate --layer mapping --phase PHS-003 --check gates

# After any change
spec-graph validate

---

Interpreting Impact Results

Key fields in impact JSON output:

{
  "affected": [
    {
      "id": "API-005",
      "type": "interface",
      "depth": 1,
      "impact": {
        "overall": "high",
        "structural": "high",
        "behavioral": "medium",
        "planning": "low"
      },
      "reason": "direct implementation"
    }
  ],
  "summary": {
    "total": 5,
    "by_type": {"interface": 2, "test": 3},
    "by_impact": {"high": 1, "medium": 2, "low": 2}
  }
}

Agent behavior rules:

• overall: high → must review and modify if needed
• overall: medium → inspect content, modify if actually affected
• overall: low → scan list only, modification rarely needed

Dimension filtering: use --dimension to focus on specific concerns

• interface change → --dimension structural
• policy/behavior change → --dimension behavioral
• schedule/scope change → --dimension planning

---

Exit Codes

Code	Meaning	Agent Action
0	success	proceed to next step
1	runtime error	check error message, retry or report
2	validation failure / gate blocked	resolve issues from output, or use --force
3	invalid input	check arguments / schema, retry

---

Caveats

• bootstrap import defaults to --mode review. Never use --mode auto.
• supersedes requires both entities to be the same type. It is directional: stored in the from entity's file. REQ-002 supersedes REQ-001 means REQ-002 is the newer entity.
• conflicts_with does not allow self-loops. It is symmetric: stored in the lexicographically smaller entity's file. Both directions are queryable via the index.
• Adding a relation that violates the allowed edge matrix fails with exit code 3. On failure, consult the edge matrix in references/data-model.md.
• metadata is a JSON string. Each type has required fields — see references/data-model.md.
• --phase is only valid with --layer mapping or --layer all. Using --phase with --layer arch or --layer exec returns an error.
• Only one plan may have active status at a time. The single_active_plan exec check enforces this.
• Entity timestamps (created_at, updated_at) are stored in TOML and populated automatically on create/update.
• After git merge with conflicts in TOML files, run spec-graph doctor to validate integrity.
• The SQLite index is rebuilt automatically on each command if TOML files changed. No manual sync needed.
• Auto-generated IDs use MAX(existing number)+1 per type. If an entity is deleted and its number was the highest, the next auto-gen may reuse that number. IDs are not stably unique across deletes. (Delete is rare and refused while relations reference the entity.)

Anti-Patterns

These are known failure modes. If you catch yourself doing any of these, stop and reconsider.

1. Mixing arch and exec concerns

Symptom: adding a requirement directly to a phase using arch-only relations, or treating a phase as an arch entity by linking it with arch-only relations. Why it's wrong: arch and exec are separate layers with separate edge matrices. Cross-layer connections belong in the mapping layer using covers and delivers. Correct approach: use covers (phase→arch) to express intent, delivers (phase→arch) to express completion.

2. Editing SQLite directly

Symptom: modifying .spec-graph/graph.db manually or treating it as the source of truth. Why it's wrong: the SQLite index is disposable and auto-rebuilt from TOML. Any manual edits are lost on the next rebuild. Correct approach: always use CLI commands to modify entities. The TOML files are the source of truth.

3. Check-driven patching

Symptom: check fails → add relations broadly until check passes → commit. Why it's wrong: passing a check does not mean the graph is correct. Over-broad relations pollute the graph and produce inaccurate impact analysis downstream. Correct approach: diagnose why the check fails, compute the minimal fix, verify semantic accuracy, then re-validate.

4. Bulk delivers expansion

Symptom: a requirement is "covered but not delivered" → add delivers for every related interface, state, and test to the phase. Why it's wrong: not all implementing entities belong to every phase. Each delivers must represent actual delivery in that specific phase. Correct approach: identify the minimal proxy set per requirement. Only entities whose delivery in this phase is necessary and sufficient to consider the requirement fulfilled.

5. Semantic ambiguity bypass

Symptom: discover a model-level conflict (e.g. edge matrix prevents a relation type the check seems to require) → work around it by expanding other relations instead of investigating the conflict. Why it's wrong: the conflict is a signal that either (a) the graph model needs revision, (b) the validator semantics need clarification, or (c) the agent's understanding is incomplete. Correct approach: when you encounter a semantic conflict between edge matrix constraints and validator expectations, stop and investigate. Check references/data-model.md for the intended semantics. If the conflict is genuine, report it to the user rather than working around it.

6. Inconsistent precision across phases

Symptom: Phase N uses broad relation additions, Phase N+1 uses precise minimal additions. Why it's wrong: the same rules must apply uniformly. If Phase 3 adds only 3 delivery proxies, Phase 2 should not have added 15 for a similar scope. Correct approach: establish the precision standard on the first phase, then apply it consistently to all subsequent phases.