v0.4.0

Changed

• The Product Spec is now agent-drafted then founder-ratified, not produced by a full interview. idea-to-release.md distinguishes the two artifacts: the Idea Brief is elicited (its intent lives only in the founder's head), while the Product Spec is derived — the agent drafts it from the approved Idea Brief and the founder documents (especially quality-bar.md), then the founder ratifies the load-bearing few (acceptance-criteria thresholds, non-goals, human-gate impacts) and the agent asks only where "good" is genuinely underdetermined, never burying a guessed threshold. AGENTS.md Feature work and the product-spec.md pointer comment updated to match. (#24)

• Two audit decisions are recorded in the maintainer docs: model-neutrality binds the operating contract (AGENTS.md, company/, templates/, docs/) while the README adapter list and historical provenance notes may name vendors (MAINTAINING.md invariant 2); and templates/experiment.md stays deliberately unwired until the 0.5.0 experiment loops (ROADMAP.md). (#60)

Added

• "Executing build tasks" discipline in company/workflows/idea-to-release.md for delegating build work to workers (a subagent or any external CLI that reads AGENTS.md): one task with tight scope, isolate per branch/worktree, least privilege (read-only for review, write only for implementation, no sandbox bypass outside a disposable runner), verify-before-merge (a worker's self-report is not evidence), and keep the human gates for irreversible or outward-facing actions. Model-neutral by design — no vendor named. (#27)
• Altitude-and-resolution guard for the Idea Brief / Product Spec interview in company/workflows/idea-to-release.md: the interviewer holds each artifact at its altitude (the founder owns intent; the agent owns mechanism), defers implementation detail the founder volunteers to the Spec and Build phases instead of deepening to match, and states the decision-grade resolution it needs so the founder neither over-specifies (anchoring the agent) nor under-specifies (confident-wrong output). Found by running the #20 interview on a real product. (#22)
• Founder-interview discipline for the Idea Brief in company/workflows/idea-to-release.md: the brief is elicited through a bounded founder interview, not drafted unilaterally, reusing the bootstrap-interview.md interviewer stance with stage-specific additions (ground ambition in feasibility, make the quality bar checkable, distill rather than transcribe). Wired from AGENTS.md Feature work and pointer comments in both templates. #20 originally produced the Product Spec through the same interview; the drafted-then-ratified model described under Changed (#24) supersedes that part. (#20)
• AGENTS.md Context discipline section: an affordance-style guide to loading context — state files as the index, current_mission / active_spec pointers for the current step, load-on-demand as the default, prefer fetching over guessing, and escalate rather than read everything to infer a missing decision. Framed explicitly as affordances and defaults, not a fixed read-list, so it orients a strong orchestrator instead of caging it while giving a smaller sub-agent a floor. (#18)
• company/workflows/founder-doc-revision.md: how to revise confirmed founder documents — everyday evolution by moving items between sections, substantive changes recorded as ADRs, with the weekly operating review as the cadence and a no-silent-change rule. Referenced from AGENTS.md and weekly-operating-review.md. (#16)

Fixed

• doctor.py fails when company/agents/ contains no role contracts: the escalation-vocabulary check silently no-opped on an empty or missing directory, so a repo with zero agent contracts passed although AGENTS.md operates through them. Fork-safe — any *.yaml satisfies it; the shipped eight are not required by name. Locked by a fixture test. (#57)
• The first weekly operating review actually gets scheduled: the bootstrap exit gate also sets next_review (workflow exit step and state-updates line, session_start.py exit-step text, state-file comment). Previously the field's only writer was the weekly review itself, so nothing ever triggered the first one. (#62)
• The release checklist specifies that the version bump in ai-company.yaml produces full SemVer (x.y.z) — releases so far wrote X.Y (0.3), leaving the field MAINTAINING calls the SemVer single source of truth in a non-SemVer format. The field itself normalizes at the next release, per the founder's decision; no mid-cycle edit. (#58)
• The Definition of Ready is satisfiable where it is checked: the Idea Brief covers the brief-level items, acceptance_criteria_testable is satisfied by the derived Product Spec, and the full gate passes before Build. Previously both workflow references told the agent the brief alone satisfies a gate whose criteria item is — post-#24 — a Spec artifact. Gate vocabulary unchanged. (#55)
• The repository map in README.md and README.ko.md matches the tree: all five workflows (founder-doc-revision was missing) and the full docs/ set (research and incidents were missing). (#53)
• The resume protocol in MAINTAINING.md no longer hardcodes a milestone that goes stale after every release: the issue-list step takes the current milestone from ROADMAP.md's "Next:" heading (it pointed at the already-released v0.3.0, so a resuming session saw zero planned issues). (#51)
• session_start.py's operating-mode instruction matches the #18 Context discipline — start from the state-file index and load what the current step needs — instead of the fixed read-list ("Read current specs, decisions, review queues, and evidence") that predates it. (#49)
• The worker-delegation topology in idea-to-release.md reads as defaults, not mandates: a single task file per worker and per-worker branch/worktree isolation are the stated defaults with their override criteria, matching the contract's affordances-and-defaults posture. The safety bullets — least privilege and no sandbox bypass, verify-before-merge, human gates — stay firm. (#47)
• The operating loop in idea-to-release.md is fully wired: Verify names definition_of_done in ai-company.yaml as the completion bar (previously the manifest's done-gate was invoked by no workflow), Observe routes failures into incident-to-learning.md, and Learn hands the next mission back to Discover (stage: discovery) — previously the loop dead-ended at learning. (#45)
• Every state field has a named writer: escalations record the blocked task's id in blocked_by_human and clear it on resolution (AGENTS.md); open ADR ids are tracked in active_decisions (docs/decisions/README.md); the weekly operating review keeps active_products and portfolio_stage current. Previously all four fields were written by nothing — company-state.yaml's header even named writers that never wrote it — and blocked_by_human's comment promised "human-assistance-request ids" although escalations only carry a task_id. Comments fixed to match. (#43)
• doctor.py requires the templates the workflows build artifacts from — idea-brief, product-spec, adr, incident, research-note — so deleting one fails the contract check instead of leaving a workflow step silently unexecutable. Locked by a new fixture test. (#41)
• templates/adr.md satisfies its governing schema and uses one id scheme: a ## Question section (decision.schema.json requires question) and Id: ADR-NNN — previously the title said ADR-NNN while the Id line said DEC-NNN. (#39)
• templates/product-spec.md can record the ratification the workflow requires: a Status: draft <!-- draft | ratified --> line and a ## Founder ratification block — one checkbox per explicitly-confirmed item from idea-to-release.md (acceptance criteria incl. agent-chosen thresholds, quality bar and non-goals, human-gate impacts) plus a ratified date — mirroring the Idea Brief's founder-decision block. Previously nothing in the artifact distinguished a draft spec from a ratified one. (#37)
• Artifact instructions name their destinations: the Idea Brief and Product Spec are written using their templates into the mission's folder under docs/specs/ — previously idea-to-release.md said "recorded in templates/idea-brief.md", which reads as writing into the blank template — and incident records get a home, the new docs/incidents/ (with README), named from incident-to-learning.md. (#35)
• The bootstrap exit gate records current_mission (the approved first Idea Brief's path) alongside stage and project_name — in the workflow's section-6 state updates and exit step 2, the session_start.py exit-step text, and the project-state.yaml field comment. Previously the brief was approved during bootstrap but nothing pointed at it, so the next session's state index (AGENTS.md Context discipline) lost the first mission. The mechanical gate itself is unchanged. (#33)
• idea-to-release.md no longer describes acceptance criteria as an interview product: the Idea Brief interview elicits the concrete quality bar, and the derived Spec restates it as observable or machine-checkable criteria — matching the drafted-then-ratified model. The interviewer stance is scoped to the brief interview plus the Spec's ratification questions, and founder-volunteered implementation detail is parked in docs/decisions/ (status deferred), the convention bootstrap-interview.md already uses. (#31)

v0.3.0

The dogfooding-hardening release: every change came from running the template on a real company.

Highlights

• Faithful capture — the bootstrap interview records founder answers in concrete detail (own words, examples, numbers), not a compressed summary.
• Rigorous interviewer — a non-sycophantic interviewer that probes vague answers and digs for intent, while the founder still owns the vision.
• No skipped questions — interview sub-questions are numbered, with a per-section coverage check before anything is written.
• Founder altitude — founder docs stay founder-level; implementation details route to docs/decisions/ (deferred).
• Research has a home — docs/research/ and templates/research-note.md, and the Customer & Market Research agent's evidence log now has a location.

See CHANGELOG.md for the full 0.2.0 → 0.3.0 diff, and ROADMAP.md for 0.4.0 ("prove it works end to end").

v0.2.0

First public release of the AI Company Starter — a git-native operating system template for a founder working with AI agent teams.

Highlights

• Reliable, documented bootstrap exit gate (confirm the five founder docs, advance the stage).
• Fixture-based tests, so CI stays green after you configure the template.
• Real contract validation in doctor.py (stage, portfolio stage, and escalation vocabulary).
• Artifact templates wired to their JSON schemas, plus evidence and learning-entry templates.
• Public-repo docs: MIT license, CONTRIBUTING, CHANGELOG, English + Korean README.
• Cross-session maintenance system (MAINTAINING.md, ROADMAP.md).

See CHANGELOG.md for the full 0.1.0 → 0.2.0 diff, and ROADMAP.md for what's next (0.3.0).

Use it: click Use this template, or gh repo create my-company --template jangjun7091/ai-company --private --clone.