Operating Model Bootstrap

When to use

• Starting a new repository or team project that will use AI agents or agent-harness skills
• Adopting this harness into an existing project for the first time
• Defining or repairing authority, risk tiers, lane isolation, or evidence-binding contracts
• Coordinating parallel human and agent lanes safely (worktrees, ownership, integration)
• Binding review to an exact candidate commit before R2/R3 work
• Correcting fragmented or contradictory rules across AGENTS.md, CLAUDE.md, and GEMINI.md

Establish a portable operating kernel and one tracked project profile. Treat humans, agents, skills, tools, and gates as one delivery system with explicit authority, ownership, evidence, and completion semantics. The day-one seed lets a new team begin system design and delivery planning while explicitly owned project unknowns are resolved.

Resolve the installation before running scripts

Keep the working directory at the consumer project root. Find this SKILL.md in the host’s selected project, user or plugin installation, and resolve scripts/ beside it. Use that absolute script path; do not assume .agents/skills/ exists in the consumer. In the commands below, replace <bootstrap-skill-dir> with that discovered directory. Never run bootstrap against the installed package or a global home directory. If more than one installation is available, use the project’s declared version/source and resolve ambiguity before mutation. Preserve existing adapters; conflicts require a deliberate merge.

Project awareness — seven areas to establish at init

The bootstrap must establish or explicitly own each of the following. Required areas must be answered before the profile leaves seed state. Optional areas have a recommended default; record not yet established — owner: role; required before: trigger for any unknown rather than leaving it blank or fabricating a value.

Backfill from an existing repo

When adopting into a repo that already has code, do not present a blank seven-area form. Run the read-only inspection pass to pre-fill what the repository already reveals:

python3 "<bootstrap-skill-dir>/scripts/inspect_repo.py" .

It reads manifests, tool configs, CI, CODEOWNERS, and docs, and prints inferred findings — each a machine guess with an evidence pointer and the exact inferred — source: <evidence>; confirm: <role> marker to paste into the profile. Rules:

• Transcribe each finding into the profile as its inferred marker; never silently promote it to a verified fact.
• Confirm inferred fields one at a time with the named human; a confirmed field becomes a plain verified value. The profile cannot go active while any inferred field remains (the validator enforces this).
• The pass is read-only and never infers authority — CODEOWNERS handles are roster candidates only; role and accountability stay unknown until a human assigns them.

Init artefacts

The initializer script seeds structure, not decisions. Running bootstrap (or the init command, which runs bootstrap first) creates these records with template content to be grounded from verified project evidence:

• docs/operating-model/PROJECT-OPERATING-PROFILE.md — the project contract, holding the seven-area answers above.
• docs/VISION.md — product vision, objectives, actors, scope, and out-of-scope (area 1).
• docs/ROADMAP.md — the big-picture scope of work: ordered outcome gates, seeded unpopulated. The team adds dates as the Product Owner confirms them.
• docs/STATUS.md — the derived situation-report structure, seeded unpopulated.
• docs/operating-model/DELIVERY-WORKFLOW.md, CHANGELOG.md, and the AGENTS.md / CLAUDE.md / GEMINI.md adapters.

The script seeds structure; the agent records the decisions. As part of the bootstrap workflow, capture the choices made during init as durable, referenceable authority:

• Baseline ADR — author the project's first architecture decision record at its ADR convention (ADR/ADR-0000-baseline-structure-DRAFT.md, approved to -approved.md; see the-architect), capturing every choice confirmed or deferred during init, with the owner and resolving trigger for each explicit unknown. This is the authority record for the starting structure.
• First status entry — add one entry to docs/STATUS.md recording the initialization event and the open unknowns. This records what happened, not invented progress; roadmap outcomes stay unpopulated until the Product Owner validates them.

Do not fabricate answers to populate any of these files. An honest unknown — owner: X; required before: Y is better than a plausible-sounding invention that will mislead every agent that later reads the profile.

Portability boundary

The operating model is model-, vendor-, and IDE-agnostic. Thin surface adapters only make the same contract discoverable and invocable. They may not change authority, risk, evidence, review, or completion semantics. This is a team-project harness bootstrap, not a production application scaffold or a substitute for project/domain controls.

Load the assets

Read the four contract assets before editing the target project:

• assets/OPERATING-MANUAL.md — released vendor-neutral normative kernel; copy it unchanged.
• assets/PROJECT-OPERATING-PROFILE.template.md — project-specific contract.
• assets/CHECKPOINT.template.yaml — durable state for R2/R3 work.
• assets/EVIDENCE-MANIFEST.template.yaml — exact-candidate evidence and review record.

Read the five planning-seed assets before initialising project delivery records:

• assets/VISION.template.md — durable product direction and initial focus.
• assets/DELIVERY-WORKFLOW.template.md — project mapping of the shared lifecycle.
• assets/ROADMAP.template.md — ordered outcome gates, initially unpopulated.
• assets/STATUS.template.md — derived situation-report structure, initially unpopulated.
• assets/CHANGELOG.template.md — project release-history structure.

Use the thin starter files under assets/adapters/ for agent surfaces the target project supports. They are discovery adapters, not competing constitutions. For a greenfield adoption, read references/DAY-ONE-EXAMPLE.md to calibrate what may start in seed state and what must wait for active controls.

Do not copy project details into the universal manual. Bind its exact SHA-256 in the profile and adapters. Unknown profile facts must be owned and time-bounded, never fabricated. Never call a seed profile active while applicable placeholders remain.

Initialise safely

From the consumer project root, using the discovered skill directory, preflight and install the day-one seed:

python3 "<bootstrap-skill-dir>/scripts/bootstrap_operating_model.py" \
  --dry-run --project-name "<project name>" .
python3 "<bootstrap-skill-dir>/scripts/bootstrap_operating_model.py" \
  --project-name "<project name>" .

The initializer creates the manual, seed profile, checkpoint/evidence templates, vision, delivery workflow, blank roadmap/status structures, project changelog, and AGENTS.md, CLAUDE.md, and GEMINI.md adapters. It preflights the whole file set, preserves different existing project planning records, and refuses conflicting operating contracts or adapters before writing anything. Map existing planning records in the profile; merge a protected contract deliberately rather than forcing replacement.

Bootstrap workflow

1. Ground the project. Read its existing constitutions, architecture, delivery flow, test commands, deployment path, protected surfaces, and current state. List contradictions and stale claims before adding policy.
2. Classify risk. Map project work into R0–R3. Round ambiguity upward. Name the highest-impact silent failure, rollback owner, and observation surface.
3. Confirm authority. Separate useful intent from permission at the classified risk tier. Obtain explicit approval before constitutional, destructive, credential, external, deployment, production-data, or capital changes.
4. Install the kernel. Run the safe initializer or copy the released manual byte for byte to the canonical governance location. Record version, digest, owner, adoption status, and review trigger.
5. Ground the profile. Resolve the day-one minimum first. Fill applicable fields with verified facts and use not applicable — reason or not yet established — owner: role; required before: trigger elsewhere. Keep status seed for design/R0/R1; require active and runnable controls before R2/R3. Never invent a command, role, or control.
6. Route baseline skills. Confirm paths or equivalents for orchestration, architecture, specification, domain/adversarial validation, candidate review, release readiness, status, and cost. Invocation syntax may vary; semantics may not.
7. Define the squad. Assign one accountable human/operator, one integration owner, bounded mutating lanes, and independent reviewers. Give every lane owned files, owned resources, and a falsifiable success test. A skill supplies capability; it never supplies authority.
8. Wire lightweight adapters. Use only the adapters the team needs. Keep the generated protected block identical across surfaces and surface-specific workflow details outside it. Run the deterministic drift check.
9. Isolate mutation. Require one branch/worktree and resource namespace per mutating lane. Read-only reviewers may share. Only the integration owner assembles pinned lane commits in a clean worktree.
10. Bind evidence. Instantiate both checkpoint and evidence-manifest assets for R2/R3. Retain exact base/candidate identity, commands, exit status, environment, data provenance, artifacts, reviewers, conditions, and rollback. Any candidate change invalidates prior review.
11. Enforce verdicts. PASS means zero conditions. Conditional/amendment verdicts block until fixed and re-reviewed; REJECT/BLOCKING halts. Make any override an explicit operator path with a durable reason, never a passing reviewer token.
12. Close the lifecycle. Run focused and convergence gates, obtain independent review of the exact integrated candidate, deliver only with authority, observe the result, and reconcile status, changelog, HLD/runbooks, evidence, open risk, and rollback.

Minimum mechanical controls

Implement controls proportionate to the project:

• a profile/adapter drift check;
• a protected-path gate that detects edits, modes, deletions, and renames;
• regression-first tests for defects;
• one convergence command for the assembled candidate;
• secret and dependency scanning;
• exact-snapshot review invalidation;
• post-delivery health/rollback checks.

Prefer a required remote check or signed attestation for high-risk approval. A local hook is defence in depth and does not authenticate reviewer identity.

Validation

Before completion:

1. Validate the distributed assets before adoption:

   python3 "<bootstrap-skill-dir>/scripts/validate_operating_model.py" \
     --template-root "<bootstrap-skill-dir>"

2. Validate an adopted seed. Before R2/R3, repeat with --require-active and pass the resolved checkpoint/evidence paths:

   python3 "<bootstrap-skill-dir>/scripts/validate_operating_model.py" \
     --target .
   python3 "<bootstrap-skill-dir>/scripts/validate_operating_model.py" \
     --target . --require-active \
     --checkpoint docs/operating-model/checkpoints/<task-id>.yaml \
     --evidence docs/operating-model/evidence/<task-id>.yaml

3. Verify all mutating lanes and external resources have one owner.
4. Change one adapter inside its protected block; confirm semantic drift fails.
5. Rename a protected file; confirm the project-specific protected-path gate catches it.
6. Change the candidate after review; confirm checkpoint/evidence binding fails or is stale.
7. Run the project's lint, type, test, security, and convergence commands.
8. Run the Adversarial Gate: name how authority, isolation, evidence, delivery, or observation could still fail silently.

Required outputs

• Adopted universal manual with version/digest.
• Completed, tracked project operating profile with all seven project-awareness areas recorded (verified fact or explicit unknown with owner and trigger).
• ADR/ADR-0000-baseline-structure-DRAFT.md (approved to -approved.md) — baseline ADR capturing all init decisions and deferred unknowns.
• Initializer-seeded docs/VISION.md, docs/ROADMAP.md, and docs/STATUS.md, grounded from evidence; the first docs/STATUS.md entry records the init event and open unknowns.
• Grounded delivery workflow plus project changelog.
• Lightweight surface adapters plus drift enforcement.
• Durable checkpoint for active R2/R3 work.
• Exact-candidate evidence manifest for active R2/R3 work.
• Validation evidence and explicitly owned unresolved platform/operator actions.
• Updated project status, changelog, and architecture/runbook documentation when the operating contract or delivery flow changes.

Do not claim complete enforcement while local checks remain bypassable, reviewer identity is unauthenticated, credentials remain exposed, or delivery has not been observed.

Release policy validation

Before release, resolve the profile's declared mode against its approved ADR and use release-contract.md for the existing validator's release-preflight invocation and evidence format. Ordinary profile validation does not establish release readiness; adopter wiring and human verification remain required.