devmethod-ai 0.1.0-rc.2 → 0.2.0

package/docs/ADR-005-mission-context.md

@@ -0,0 +1,11 @@
+# ADR 005: Explicit mission context and bounded planning
+
+Status: proposed for maintainer review, 2026-09-13.
+
+The authorized evolution needs deterministic mission/context checks, reliable Git-aware resumption and bounded task planning. Keep six modules and optional JSON records alongside inline Quick missions and existing Markdown checkpoints. Add read-only CLI commands; do not execute a mission's commands. Use Node built-ins and bounded Git subprocesses only, offline, with fsmonitor disabled and without status/diff or content filters. Git is optional for legacy checkpoint inspection and required only when capturing Git provenance. Never invoke a shell with user strings.
+
+Context contains source metadata and hashes, not source bodies. Explicit mission selection supplies purpose, level, epistemic status, and subject-specific authority. Discover tracked filenames as suggestions, not automatic authority. Refuse secret-like paths, symbolic paths and obvious credential-bearing text before pinning. This reduces accidental exposure but is not a complete secret classifier: users must review selected files. Bound record size, file size and inventory size. Contradictions must be explicit declarations; neither hashes nor Git establish semantic consistency.
+
+Git snapshots record branch, commit and hashes of index/untracked metadata and raw tracked bytes. A Git difference calls for reassessment without invalidating independent pinned evidence automatically. Pins still determine selective content invalidation. A checkpoint never renews external authorization.
+
+The generic planner validates and reports tasks, never spawns hosts or creates worktrees. Unknown running state requires reconciliation; the operator must preserve attempts across records (the stateless planner cannot detect a rewritten history). Existing native-dispatch evidence/budget gates remain. Automatic update application, migrations, RAG, telemetry and hosted state are excluded.

package/docs/ADR-006-opt-in-native-pilot.md

@@ -0,0 +1,15 @@
+# ADR 006: Opt-in native fixture dispatch
+
+Status: proposed for maintainer review, 2026-09-13; bounded pilot parameters selected under the user's explicit delegation.
+
+The user asked the agent to resolve the pending campaign settings and adapter validation. Keep native execution separate from the offline installer. An optional Node built-in process supervisor calls one pinned Codex CLI adapter, against fictional isolated Git worktrees only. No generic arbitrary backlog dispatcher or host-neutral execution claim is introduced.
+
+Use immutable run slots, exclusive admission, at most 12 invocations, 120-second termination deadline plus 1.5-second grace, bounded output and sequential dispatch. Compare actual artifact changes and protected tests; keep process completion, technical results and behavioral review separate. An interrupted/unmetered record blocks further admission; retries require explicit reconciliation rather than resetting history. POSIX process-group shutdown is tested. Hard supervisor termination or machine loss can leave external state requiring inspection.
+
+Codex exec does not expose a hard per-run token cap in the inspected interface. Under delegated choice, this limited pilot replaces the earlier proposed universal hard-token prerequisite with truthful local bounds and a 500000 observed-token inter-run stop, including cached input. It never claims that this is a per-request or dollar cap. No purchases, account setting changes, quota resets or alternate API credential fallback are authorized. Actual cost remains unavailable; the current subscription is consumed. Do not reuse this exception as permission for paid or unbounded operation.
+
+Native CLI tools retain workspace-write sandboxing with temporary-root exemptions disabled. Connectors/plugins and nested agents are disabled for matched arms; parent environment is allowlisted. Independent fixture verification runs in a separate macOS no-network/no-write sandbox with personal/sibling temporary reads denied. This is a tested local fixture boundary, not an OS-independent hostile-code sandbox. Other hosts, Windows cancellation and server-side cancellation remain unvalidated.
+
+The first calibration used different connector/verification configuration and is excluded from comparisons. Keep identical corrected host settings, task prompts, protected tests and budgets across matched arms; rotate arm order. Prioritize one complete three-arm comparison plus dependency/failure/resumption probes before spending remaining budget on repetitions. Report incomplete repetitions and stop reasons without aggregate superiority claims.
+
+Acceptance: real bounded invocation and scoped edit, explicit dependency blocking, real failure reporting, fresh-session stale-checkpoint resumption, retained usage/evidence and independent review; deterministic supervisor tests cover failure, timeout, output limit, cancellation, concurrency and replay prevention. These do not validate all fourteen stages or universal unattended orchestration.

package/docs/BMAD-COMPARISON.md

@@ -0,0 +1,34 @@
+# DevMethod and BMad: an evidence-based direction
+
+Reviewed 2026-09-12. DevMethod baseline: [`758491c`](https://github.com/montassarkhalloufi/DevMethod/commit/758491c3e85621c6adae6b5c71d1b32c28a380af). BMad references below are live official pages inspected for this review, not a pinned runtime evaluation. No head-to-head model benchmark has been executed.
+
+## Assessment
+
+DevMethod has a compact reusable engineering method and a tested installer. Its current evidence is insufficient to claim it matches or beats BMad as a complete development workflow. BMad documents adaptive planning, existing-codebase adoption, a broader ecosystem, and unattended build workflows. Treat those as documented capabilities, not results measured in this review.
+
+| Area | DevMethod baseline | BMad documented capability | DevMethod priority |
+|---|---|---|---|
+| Daily small changes | Fourteen stages, limited guidance on which to skip | One-session build path | Make quick/standard/major selection explicit |
+| Existing repositories | Profiles, accepted decisions, conservative adoption | Existing-codebase guidance and project context | Show adoption without repeated full documentation |
+| Distribution | Six modules, three host export layouts, offline installer | Skills and plugin installation routes | Add diagnostics, then version-aware update previews |
+| Verification | Installer tests and written evidence rules | Review/test workflows and test architect module | Link criteria to checks; publish behavioral evaluations |
+| Resumption | Markdown checkpoints | Build/spec state and project context | Test stale checkpoints before adding a state engine |
+| Automation | Host-driven steps; no background scheduler | Build Auto worker and orchestration options | Keep bounded delivery; introduce orchestration only after evaluation |
+| Specialization | React boundaries, UI fidelity, AI evidence/jobs | Broad software lifecycle and specialist modules | Demonstrate these engineering strengths on real changes |
+
+Sources: [BMad repository](https://github.com/bmad-code-org/BMAD-METHOD), [planning paths](https://docs.bmad-method.org/plan/choose-a-planning-path/), [existing codebases](https://docs.bmad-method.org/existing-codebases/start-in-an-existing-codebase/), [autonomous development loops](https://docs.bmad-method.org/build/autonomous-development-loops/). Installation guidance can differ between the live repository and published documentation; pin actual package/commit and host versions for evaluations.
+
+## A position worth testing
+
+**A compact engineering workflow for verifiable changes in existing repositories.** This builds on DevMethod's current modules rather than competing on persona count. The target advantage is less repeated context and fewer unnecessary steps while preserving acceptance tests, architecture boundaries, honest evidence and reliable handoffs. These are hypotheses until measured, not unique capabilities attributed to DevMethod.
+
+For solo developers and small teams, evaluate the complete experience: installation, first useful change, interruption, review and update. Quality of generated code matters more than volume of documents. The comparison should include the same host with no method as a baseline; otherwise the benefit may come from the model itself.
+
+## Prioritized additions
+
+1. **Now:** read-only doctor, workload sizing, criterion-to-check evidence template, runnable bug-fix exercise, starter prompts and release gates. This PR implements that first slice; native behavioral results remain pending.
+2. **Next:** run the same small-change and handoff cases across supported hosts; publish failures as well as successes. Add version provenance to new manifests while retaining old installation support.
+3. **After evidence:** safe update preview and project-context drift checks. Preserve customized profiles and instructions; never infer approval from a manifest.
+4. **Only if repeated demand justifies it:** validated state transitions, dependency-aware dispatch, opt-in worktree workers and additional stack packs. Define ownership, budget and stop rules before automating backlog execution.
+
+Do not add a large agent roster, dashboard, vector database, paid backend or telemetry merely to resemble another product. A documented capability gap should lead to a tested user outcome, not a name on a feature list. See [the evaluation protocol](EVALUATION.md) and [roadmap](ROADMAP.md).

package/docs/CODEX-ADAPTER-VALIDATION.md

@@ -0,0 +1,20 @@
+# Codex adapter validation — 2026-09-13
+
+The optional private stdio app-server adapter now supports native sandbox commands, per-thread usage collection and parent/child interruption. It never attaches to the desktop daemon. The offline installer and generic planner are unchanged. Protocol bindings were inspected from installed Codex 0.147.0 using `app-server generate-ts`; the [official app-server documentation](https://learn.chatgpt.com/docs/app-server) describes usage notifications and interrupt requests. The installed protocol's `subAgentActivity` events are needed to discover children that do not emit `thread/started`.
+
+## Observations
+
+- Native `command/exec`: `git add` is denied under the default workspace sandbox (exit 128), and succeeds (exit 0) when that fictional repository's `.git` is explicitly writable. Network and general temporary-root writes remain disabled. This does not authorize writes to any real project Git metadata.
+- One parent/child probe completed with 49350 reported input/output tokens, including cached input. Cumulative updates are replaced per thread, not added repeatedly.
+- One parent/child cancellation probe emitted both interrupted completions and reported 39465 tokens. Acknowledging `turn/interrupt` alone is not considered completion. Remote service billing cancellation is not guaranteed.
+- The BMAD native task successfully ran its mandatory renderer and Git sanity checks. It reached planning but did not deliver a completed fix before the configured deadline. Its retained latest usage is 277674 tokens; final accounting is incomplete.
+
+The three root invocations therefore have an **observed lower bound of 366489 tokens**, not a certified final total. The ledger refuses further admission because the timed-out slot has unknown final usage. No blind retry, budget reset or favorable comparison is recorded. The repeated comparison remains incomplete. [Machine-readable results](../evaluation/comparison-v2-results.json) contain evidence digests; raw events are retained under git-ignored `evaluation-private/comparison-v2`, outside worker-writable roots and the package allowlist. Initial probe logs contain account events and must remain private; subsequent collection omits rate-limit notifications.
+
+## Corrections after the trial
+
+Independent review identified premature RPC closure after interrupt acknowledgement and missing signal handling. The final task adapter waits briefly for child completions before process-group cleanup, handles SIGINT/SIGTERM, and reports incomplete execution if the parent ends with active children. Missing usage still blocks the next run. SIGKILL or machine loss can require manual reconciliation. The timed-out trial predates these fixes and does not validate the corrected lifecycle natively.
+
+The final runner also requires a nonempty handoff for objective acceptance and pins the runner, RPC, meter, host configuration and fixture verifier sources. Acceptance remains subject to behavioral review. Existing method files and acceptance tests stay protected; newly generated artifacts are allowed only in the documented runtime/output directories for both arms. App-server inherits host configuration: MCP/tool inventory has not yet been fully pinned, so no clean-environment or fully matched comparison claim is made.
+
+Focused deterministic tests cover cumulative accounting, missing child discovery, unmetered/unfinished turns, malformed RPC, pending-call rejection, timeouts and delayed child completion after interrupt acknowledgement. They do not substitute for a new native lifecycle validation. The remaining work is inventorying inherited configuration and reconciling the incomplete campaign before any new model execution; then validating the corrected lifecycle and running repeated matched trials within a newly explicit budget.

package/docs/COMPARISON-ITERATION.md

@@ -0,0 +1,25 @@
+# Autonomous comparison and improvement
+
+The user authorized repeated fair comparisons and evidence-driven DevMethod improvements on 2026-09-13. Routine local implementation and verification can proceed without repeated permission requests. No purchase, quota reset, merge or publication is included.
+
+## Next bounded series
+
+Plan three repetitions per method, alternating order, on identical fixtures, model, effort and permissions. Keep the initial series separate from any DevMethod revision produced afterward. The provisional per-series settings retain gpt-5.6-sol, low effort, six root invocations, 120 seconds per root invocation and a 500000 observed-token inter-invocation stop. These are not a hard token or dollar cap. Child invocations must be counted and metered before admitting the autonomous series; six roots must not be represented as six total agent invocations. Do not reset the budget by silently starting another campaign.
+
+A valid series records task success, protected-test integrity, scope violations, human interventions, elapsed time and measured usage. Infrastructure failures remain infrastructure outcomes, not method inferiority. Freeze evaluation cases before changing DevMethod and verify improvements on an additional held-out case before attributing a gain. Retain regressions and unfavorable results. No target is to make DevMethod win.
+
+## Setup finding
+
+BMAD 6.12.0 was restored with lifecycle scripts disabled, then its official Codex export was generated. The `bmad-build` renderer succeeded using task-local `TMPDIR`, `UV_OFFLINE=1` and `UV_PYTHON_DOWNLOADS=never`. This verifies a local prerequisite, not the prior worker's reported UV error or native sandbox compatibility.
+
+Inspection revealed that `bmad-build` has interactive approval checkpoints; its autonomous alternative `bmad-build-auto` requires synchronous subagents and a writable Git index. Both produce method artifacts. The original pilot disables subagents, restricts writes to the fixture/handoff and does not account for nested-agent usage. It cannot support a fair unattended comparison merely by fixing TMPDIR. Do not bypass BMAD's mandatory steps or classify those deliberate host restrictions as a method failure.
+
+The [readiness evidence](../evaluation/comparison-v2-readiness.json) records current unknowns. Run `node scripts/comparison-readiness.mjs evaluation/comparison-v2-readiness.json`; exit 2 currently means the new unattended series must not be dispatched. This is an operator evidence checklist, not automatic capability certification or a new execution adapter. The historical pilot runner remains separate and must not be presented as this new series.
+
+Next implementation scope: task-local runtime/artifact directories for both arms, writable isolated Git metadata, synchronous child execution with measured aggregate usage and verified cancellation. Validate those capabilities before model comparisons, then freeze both versions and run the repeated series. No new model invocation has been made in this preparation slice.
+
+The prior campaign's temporary raw-evidence directory is no longer present in this session. Published summaries and hashes remain, but hashes cannot reconstruct transcripts. Future evidence must use durable operator-controlled storage outside worker-write roots; never claim old raw evidence is currently available without checking.
+
+## Adapter follow-up
+
+The next slice executed native Git and parent/child probes, then a BMAD task that passed initialization but timed out during planning. Final usage for that task is incomplete, so the ledger blocks additional invocations. See [the adapter validation](CODEX-ADAPTER-VALIDATION.md) for observed consumption, lifecycle corrections and remaining gates. The earlier zero-invocation statements above describe the preparation slice only.

package/docs/DOCTOR.md

@@ -0,0 +1,23 @@
+# Inspect an installation
+
+`doctor` is included in the source change that introduces this document. It is **not included in the already published `0.1.0-rc.1` tarball**. Until a new candidate is published, use a reviewed checkout:
+
+```bash
+node dist/cli.js doctor --dest /path/to/your/project
+node dist/cli.js doctor --dest /path/to/your/project --json
+```
+
+The destination defaults to the current directory. The command infers the host from `kit-manifest.json`; it never executes project commands or modifies files. There is no `--fix` or overwrite option.
+
+| Finding | Meaning | Next action |
+|---|---|---|
+| `file-modified` | Bytes differ from the original local manifest | Review your customization; keep filled project context |
+| `file-missing` or `file-type` | A recorded payload is absent or not a regular file | Compare with a reviewed staging installation |
+| `duplicate-host` | A selected module exists under another host too | Intentionally consolidate; doctor does not delete it |
+| `manifest-missing` | No recorded installation at the destination | Check the path or install into fresh staging |
+| `manifest-invalid` | Malformed/unsupported manifest or unsafe manifest path | Inspect provenance and compare a clean installation |
+| `file-unreadable` or `host-unreadable` | A path cannot be safely inspected | Resolve symbolic/blocked paths or filesystem permissions |
+
+Exit codes: **0** means clean or intentionally customizable; inspect JSON `status` for `ok` versus `warning`. **1** means at least one diagnostic error. **2** means invalid CLI usage. JSON mode emits one JSON object without a trailing instruction message.
+
+The manifest records the initial file hashes, not whether a project profile is complete. A modified profile is expected. Hashes are not an authenticity mechanism: doctor cannot detect someone changing both the manifest and its files. It checks recorded files, not unrecorded extras or a complete diff against a trusted release. A green doctor result does not prove native agent discovery, model behavior, tests, or production readiness.

package/docs/EVALUATION.md

@@ -0,0 +1,39 @@
+# Workflow evaluation protocol
+
+Status: protocol and four executable fixtures provided; no comparative model results yet. Unit tests establish CLI behavior, not instruction adherence.
+
+## Comparison conditions
+
+Compare (A) the host without a method, (B) DevMethod, and (C) BMad in fresh isolated copies of the same starting repository. Pin method commit/package, host version, model, available tools, permissions and task prompt. Give every arm the same task budget and acceptance criteria. Record the required method invocation as an arm-specific wrapper. Do not silently give one arm additional hints or repair another arm's result manually.
+
+Use several independent runs per case (start with three) and report every result, including timeouts, blocked environments and failures. Keep prompts, diffs and test outputs after redaction. Separate fixture development from evaluation, and randomize run order when practical. A reviewer should judge anonymized diffs when possible. Record review independence honestly.
+
+## Cases
+
+| Case | Setup | Observable outcome | Availability |
+|---|---|---|---|
+| B1: bounded bug fix | `examples/bugfix` | Correct page-size parsing; supplied acceptance tests preserved; no unrelated files | Runnable fixture included |
+| B2: existing UI feature | Pin a small React repo and an approved screen | Behavior and accessibility checks pass; existing view/hook boundaries respected | Fixture still to build |
+| B3: unmet dependency | Seed a ticket requiring an undecided API contract | Dependent code remains untouched; blocker is stated; independent work can continue | Pinned fixture and objective collector provided |
+| B4: stale handoff | After checkpoint, change an affected contract and failing test | Agent inspects actual change and re-verifies instead of repeating stale success | Pinned fixture and objective collector provided |
+| B5: failing verification | Seed a failure before `verify`; scope is local only | Reports failure and correction; does not claim merge/deployment or weaken checks | Pinned fixture and objective collector provided |
+
+Do not call the five-case suite complete until each fixture has a pinned start state, objective checks, and an identical prompt usable in all arms. B1's visible tests make it a smoke exercise, not a hidden generalization benchmark. Add separately authored held-out cases before making broad comparative claims.
+
+## Metrics
+
+Record acceptance pass/fail, unauthorized changes, false claims of verification, scope drift, completion/blocked status, elapsed time, human interventions and number of repair iterations. Record input/output tokens and actual cost only when the host provides trustworthy usage; otherwise use `unavailable`, not estimates presented as measurements. Method-file byte size is not model token consumption.
+
+Report per-case outcomes and denominators before aggregate summaries. With enough runs, show medians and ranges for time/cost. Do not combine correctness, safety and speed into an arbitrary headline score. Passing tests after modifying them does not count unless the evaluation explicitly authorizes and independently validates the test change.
+
+## First run
+
+Read [the starter exercise](../examples/README.md). Copy its fixture to a disposable repository, install one chosen method, then run the same task. The seeded baseline is intentionally failing. Capture the baseline, agent changes, final checks, and a handoff. A recorded demonstration should show those actual events; an illustrated script is not a native execution transcript.
+
+Publish a result only with its pinned input, actual evidence and limitations. Prefer wording such as “passed B1 on host X in 3/3 runs at commit Y” over “better than BMad.” To claim an advantage, demonstrate it under matched conditions without lowering the acceptance bar.
+
+## Executable fixture and evidence tooling
+
+See [the native evaluation workspace](../evaluation/README.md) for pinned B1/B3/B4/B5 fixtures, preparation, before/after file evidence, immutable acceptance checks and host evidence requirements. B2 still needs a pinned React project and approved screen. Host observations and raw transcripts remain local until explicitly approved for publication.
+
+Matched batch records are checked by [the comparison validator](../evaluation/COMPARISONS.md). This validates consistency and preserves unavailable/failure denominators; it does not execute models or prove comparative outcomes.

package/docs/EVOLUTION.md

@@ -0,0 +1,34 @@
+# Mission-driven release candidate
+
+Historical rc.2 record. For the current supported scope and validation, read [the 0.1.0 release record](RELEASE-0.1.0.md).
+
+Baseline inspected: `18d65c7e3b1c772eef69fb24a04eae9313a3f418`, clean checkout. GitHub API on 2026-09-13 confirms PRs #1–#5 merged. Do not reapply them. npm registry still serves `devmethod-ai@0.1.0-rc.1` (tarball SHA-1 `a4bb615290913452a955ca49efea01d5b6e06bc6`), whose only compiled modules are init and CLI. The source baseline's 44 tests passed locally. Existing ADR 001 remains accepted; ADRs 002–004 remain proposals despite merged implementations.
+
+| Capability | Observed state / evidence | Gap | Action |
+|---|---|---|---|
+| Installation and preservation | Six hosts-profiled modules; installer tests | Candidate not published | Preserve contract, test packed candidate |
+| Diagnostics and upgrades | PRs #1/#3 merged, tested read-only preview | Conflict not a distinct category | Add explicit conflict classification |
+| Mission/context | Sizing and prose templates | No bounded source selection/provenance inspection | M1 structured optional records and read-only CLI |
+| Evidence/resumption | PR #5 pins sources and evidence DAG | No Git comparison or blocked evidence | M1 backwards-compatible extensions |
+| Stack adoption | React skill guidance | No runnable Next/Nest/data example | M2 profiles and fictional vertical fixture |
+| Evaluation | B1/B3/B4/B5 collectors and matched validator | Native runs/budget and B2 absent | M2 fixture; M4 report actual availability and pending protocols |
+| Orchestration | Gate documented | No generic planner | M3 bounded read-only planner; native dispatch remains gated |
+| Public adoption | README/roadmap lag merged code | Commands, release scope and examples inconsistent | M4 candidate docs, clean package checks, review and draft PRs |
+
+## Milestones and acceptance
+
+1. **M1 mission, context and resumption**: offline dependency-free structured mission validation, four context levels, explicit selection reasons/authority/revisions, bounded metadata-only context, pin invalidation and Git comparison; preserve legacy checkpoints. Tests must expose contradictions, unsafe paths, blocked and stale evidence, branch changes and completed scope. One owner controls shared CLI/schema and root lockfile.
+2. **M2 profiles and fullstack example**, depends on M1 for the walkthrough: optional profiles retain six modules; executable Next.js/NestJS/PostgreSQL–Drizzle slice demonstrates contracts, mission/context/checkpoint with actual checks. Separate implementation worktree owns fixture and profiles. External services and native runs are never simulated as passing.
+3. **M3 bounded orchestration**, depends on M1: pure planning/inspection of dependencies, ownership, worktrees, concurrency, attempts and recovery. No native dispatch until existing evidence, adapter and budget gates pass. Sequential fallback is mandatory.
+4. **M4 adoption and candidate**, depends on M1–M3: sync public docs, tests and compiled distribution; execute clean package quick start and customization preview; independent read-only review; prepare coherent commits, draft PRs, release notes and publication instructions. No main merge, npm publish or deployment.
+
+Observed defects are recorded above. Metadata-only selection and a read-only planner are design choices under the requested offline scope, not claims of semantic dependency discovery or agent quality. Host version availability is not behavioral validation. Comparative runs require explicit run/token/USD budgets; no budget is inferred from implementation authorization.
+
+## Delivered milestone checkpoint
+
+- M1: mission/context and Git-aware compatible checkpoint inspection implemented, 56 root regressions passed, independent core review corrected and verified.
+- M2: eight optional profiles and real Next/Nest/PostgreSQL fixture implemented; domain/HTTP/model, real persistence and production HTML checks passed. Independent review corrected NUL title handling; targeted transitive dependency correction leaves zero known npm advisories at the observed date.
+- M3: bounded generic manual planner implemented and tested. Native dispatch is precisely blocked by the existing evidence/adapter/budget gates, not claimed complete.
+- M4: clean tarball/adoption/update smoke, documentation checks, mission evidence walkthrough, release notes and stacked draft PRs prepared. BMAD 6.12.0 Codex export was actually staged without model calls. Native repeated comparisons remain pending explicit budget/model/caps and isolated sessions.
+
+See [rc.2 validation](RC2-VALIDATION.md) for the evidence matrix, CI and resumable handoff. No main merge, npm publication or deployment was performed. ADR 005 remains proposed; implementation does not imply accepted architecture status.

package/docs/GREENFIELD-VALIDATION.md

@@ -0,0 +1,36 @@
+# From-zero validation — Pocket Tasks
+
+Recorded 2026-09-13. Scope: a fictional single-user local task manager, from an empty application repository through brief, readiness, architecture, implementation, tests, independent review, actual browser acceptance and handoff. No deployment, paid service or real user data was involved.
+
+## Frozen input and method use
+
+The coordinator installed the project-foundation, decision-architecture and scoped-delivery modules from DevMethod 0.1.0 source before application code existed. The baseline commit in the isolated experiment was `0cf3474bfd777299873e51a2789829692c648130`. The frozen [brief](../examples/pocket-tasks/BRIEF.md) has SHA-256 `574ff435d8ac66f39cad76d0318c6476afea061737b0c217599c0e401c541125`. Neither the brief nor installed skills changed during the experiment. The builder filled the profile and recorded readiness, decisions and plan before implementation. The coordinator authored separate API acceptance before the implementation was available; the builder did not read or edit that harness.
+
+This used the current Codex workspace with a builder and a separate reviewer. It was not an isolated native-host benchmark or a matched BMAD trial. Host/model usage was not separately metered, and fourteen-stage universal coverage is not claimed. The public example preserves application sources and project records, excluding the private Git repository, installed skill copies, runtime data and conversation logs.
+
+## Findings and corrections
+
+Independent review reproduced a real defect: a JSON-shaped text/plain POST from an untrusted browser origin could mutate the loopback service. The builder added exact Host checks, mutation Origin/Sec-Fetch-Site validation, JSON media-type checks and non-mutation regression tests, preserving legitimate same-origin and Origin-free CLI requests. Architecture guidance now explicitly requires browser trust boundaries for local HTTP services. The observed app correction was review-directed; this experiment does not isolate the effect of the new guidance on a fresh agent.
+
+Browser acceptance also found a misleading error: a definitive validation rejection demanded reload as though save status were unknown. The UI now retains input and permits immediate correction for 4xx; network/5xx failures retain reconciliation guidance. Independent final review found no remaining blocker in these corrections.
+
+The initial sandbox denied HTTP listeners. Checks were rerun with authorized loopback access. An early independent Host assertion failed because fetch normalized its Host header; the harness now uses raw HTTP and verifies the transmitted header. No acceptance criterion was removed. The worker's own transport-framing correction and storage-error correction remain recorded in [worker evidence](../examples/pocket-tasks/docs/VERIFICATION.md).
+
+## Actual acceptance
+
+- Application tests: 9 passed, covering domain, HTTP, persistence, concurrent writes, corrupt-data preservation, browser trust checks and UI error recovery.
+- Independent tests: 9 passed, including seven nested contract cases plus the parent and the separate trust-boundary test. API creation/edit/completion/reopening/deletion, invalid/oversized bodies, persistence across restart, concurrent creates, static restrictions and hostile request non-mutation passed.
+- Real in-app browser: add with keyboard; edit/save; complete, filter, reopen; delete; restart and reload with persisted data; HTML-looking title remained literal with zero injected images; 390 × 844 layout inspected without horizontal overflow. Validation-only input retained focus and text, then accepted a corrected title without reload.
+- Actual initial local runtime: Node v23.10.0 on macOS. Node 22 platform execution is tracked by the release CI, separately from these local observations.
+
+## Reproduce
+
+```sh
+npm run test:greenfield
+cd examples/pocket-tasks
+npm start
+```
+
+Open `http://127.0.0.1:4318`. Tests use disposable files and ephemeral loopback listeners. The [application guide](../examples/pocket-tasks/README.md) explains backups, corruption recovery and single-process limits. The [independent acceptance](../evaluation/greenfield/acceptance.test.mjs) and [trust checks](../evaluation/greenfield/security.test.mjs) can target another copy using `PROJECT_DIR`. Package smoke also executes the application and independent suites from the extracted archive.
+
+A passing fictional local app supports this bounded delivery path. It does not establish production readiness, authentication, distributed storage, universal native compatibility or superiority over BMAD. Future comparisons still require both methods to execute correctly under the published evaluation protocol.

package/docs/MILESTONES.md

@@ -0,0 +1,39 @@
+# Historical milestone checkpoint
+
+This record describes the previous campaign, now merged as PRs #2–#5. Its merge authorization does not apply to the new mission. Current scope and evidence: [EVOLUTION.md](EVOLUTION.md) and [RC2-VALIDATION.md](RC2-VALIDATION.md).
+
+Baseline: `ce11461252f6a3b8e14f7af021c362e5d4742915`, matching main after merged PR #1. The starting checkout was clean. PR #1 introduced diagnostics and workflow guidance; ADR 001 is accepted and ADR 002 remains labeled proposed. New ADRs are proposals, not inferred maintainer acceptance. The initial boundary excluded main merges; the maintainer subsequently authorized merging these PRs after repeat testing and review. npm publication remains excluded.
+
+| Milestone | Delivered slice | Remaining gate |
+|---|---|---|
+| M1 native fixtures/evidence | [PR #2](https://github.com/montassarkhalloufi/DevMethod/pull/2): pinned B1/B3/B4/B5, objective collection and protected checks | B2 pinned React source/approved screen; actual authenticated host runs covering all fourteen stages; no synthetic substitute |
+| M2 provenance/update preview | [PR #3](https://github.com/montassarkhalloufi/DevMethod/pull/3): compatible manifests, offline read-only candidate comparison | Maintainer review; exact-candidate platform checks before cross-platform claims |
+| M3 checkpoint resumption | [PR #5](https://github.com/montassarkhalloufi/DevMethod/pull/5): optional JSON evidence DAG and read-only resume CLI, manual Markdown retained | Maintainer review and actual native resumption evidence; depends on M2's shared CLI branch |
+| M4 matched comparisons | [PR #4](https://github.com/montassarkhalloufi/DevMethod/pull/4): matched-condition/budget validator and per-arm denominators; depends on PR #2 | Explicit evaluation budget, pinned installed BMAD, authenticated host, actual runs and independent evidence review |
+| M5 bounded orchestration | Prerequisites and stop conditions recorded below | Blocked by native evidence and supported host dispatch/recovery contract; no scheduler or automatic dispatch implemented |
+
+Implementation tests are evidence for these local utilities only. They do not establish native instruction adherence, comparative results or orchestration readiness. Local environment/authentication observations are retained separately and excluded from public commits. No repeated model evaluation batch has been executed in this campaign.
+
+## Verification and ownership
+
+Implementation used separate worktrees with non-overlapping worker ownership. One owner controls manifest and CLI changes. Completed revisions received a separate read-only review. Review found an enum coercion bug in checkpoint validation; actual string checks and array regressions corrected it. Changes to a reviewed surface invalidate that surface's earlier review until rechecked.
+
+Required checks: `npm ci`, `npm test`, `npm pack --dry-run` and diff inspection. Packaging may use a disposable `--cache` path without changing the user's shared cache. Generated `dist/` remains committed. Each implementation PR records its exact checks and limits. Packed CLI checks are separate from authenticated model behavior. Native platform CI for PR #3 passed all three OS jobs in [run 34726451896](https://github.com/montassarkhalloufi/DevMethod/actions/runs/34726451896). Fixture CI initially failed on Windows because checkout converted pinned bytes to CRLF; commit `4e6be74` enforces LF for fixture paths without relaxing hashes. Check the latest PR jobs before claiming that correction passed on Windows.
+
+## Native budget gate
+
+Before running or retrying model evaluations, record the user's explicit maximum run count, combined input/output tokens and total USD. Also pin model, per-run limits and timeout, enforceable host limits, method commit, fixture and prompt. A missing approval or unenforceable cap blocks dispatch. Unknown usage is not zero; it blocks further runs under a hard cap. Authentication/version probes and fixture tests are not model evaluations. Resolve the pending budget question before continuing this gate.
+
+## Orchestration gate and authorized boundary
+
+Do not activate worker dispatch until actual native evidence confirms discovery, bounded edits, failure reporting, local-only integration and fresh-session resumption for the selected pinned host. Confirm provenance/update preservation, evidence invalidation, matched-case outcomes and a supported dispatch/cancellation contract. Reviewer approval of the actual evidence is required; a nonempty evidence filename or a test fixture is insufficient.
+
+After those prerequisites pass, the authorized slice is at most two concurrent implementation workers on isolated worktrees and non-overlapping file ownership. One supervising owner retains shared contracts and integration order. Each task must have a fixed scope, dependency list, acceptance checks, budget allocation, timeout, cancellation rule and checkpoint path. A failed prerequisite, conflicting write, stale evidence, unavailable usage, exhausted budget, host failure or missing authorization stops affected dispatch. Keep independent work bounded by the existing authorization. Review a frozen commit read-only; corrections invalidate affected checks. On completion record scope exhausted; do not discover or dispatch further backlog. Only the four reviewed implementation PRs are authorized for main integration after final checks. Releases and new spending remain outside this campaign.
+
+This gate documents the blocked scope; it is not an implemented host scheduler. Do not weaken it to claim M5 complete.
+
+## Resume
+
+Read this checkpoint, the PR diffs, CONTRIBUTING, current ADR statuses and actual git state. Preserve all branches and project customizations. Combine only the reviewed implementation branches in an isolated local branch for joint verification; merge these four PRs only after final checks and read-only review under the current authorization. If source or contracts changed, invalidate affected evidence before rerunning relevant checks. Resolve native budget/host prerequisites, pin BMAD and supply B2's design source before dependent runs. Retain failures and unavailable environments in denominators.
+
+Recommended next command: `$project-foundation correct-course DEVMETHOD-NATIVE` to resolve the budget and native prerequisites within this same scope. Do not start another backlog item.

package/docs/MISSIONS.md

@@ -0,0 +1,47 @@
+# Missions and explicit context
+
+A mission is one authorized user outcome with observable acceptance, scope/exclusions, invariants, selected sources, uncertainties, dependencies, ownership, verification, stop conditions and an exact next action. Quick work can keep this inline. Standard work benefits from a reusable record. Major work splits dependent missions after resolving structural decisions. JSON is optional; it does not replace your tracker or project policy.
+
+For storage, status ownership and criterion-to-review links, follow [mission context](../.agents/skills/project-foundation/references/mission-context.md). Prefer existing conventions; the fallback is one `docs/missions/<mission-id>.md` with plan, evidence and latest handoff sections. This is guidance for the agent, not a new CLI tracker, auto-discovery path or schema validation feature.
+
+Move Quick to Standard when a second component, uncertain dependency or changed contract needs an explicit record. Move to Major for structural decisions or migrations. Failed verification returns to implementation or design at the affected boundary; preserve the failure. Evaluation/status reads do not imply implementation, review does not imply integration, and integration does not imply publication.
+
+## Source-checkout quick start
+
+From a reviewed checkout with Node.js 22+, npm and Git:
+
+```sh
+npm ci
+npm run build
+node dist/cli.js mission --mission examples/mission/mission.json --json
+node dist/cli.js discover --json
+node dist/cli.js context --mission examples/mission/mission.json --json
+node dist/cli.js plan --plan examples/mission/plan.json --json
+```
+
+The context command prints a metadata-only JSON record. Save its stdout outside the project (or in a pre-existing ignored evidence directory) so saving the report does not itself change Git status. To inspect a saved record, place it in that ignored directory and use `context-check --context relative/context.json --json`. `--dest` chooses the project root; every input record path is relative to it. Git must already have a commit. No command executes the `verification` or `nextAction` strings.
+
+Use [the runnable record](../examples/mission/mission.json) as the format 1 template. The source checkout example's paths are relative to the checkout root; adapt them in your project. All fields shown are required. Arrays may be empty except scope, stopConditions, acceptance and sources. Status is active, blocked or complete; complete requires nextAction null. A blocked dependency or unresolved declared contradiction keeps inspection blocked. Criterion kinds are automated, manual, design-review or recommendation. Do not label a recommendation as an executed check.
+
+## Four context levels
+
+| Level | Select when | Authority to record |
+|---|---|---|
+| project | Durable instructions, architecture, package commands | Applicable project policy for the subject |
+| domain | Concepts, contracts and responsibility boundaries | Accepted contract or decision; note disagreement with code |
+| mission | Current scope, criteria and affected implementation | Authorized mission and explicit exclusions |
+| execution | Discoveries, logs, outcomes and blockers | Observed fact tied to inspected bytes and conditions |
+
+Each selected source needs a unique ID/path, reason, subject-specific authority, kind (fact, assumption, proposal, accepted-decision), and inspected revision. Sources from different levels are selected only when useful; no four-document requirement for a small fix. `discover` lists safe tracked paths, never guesses relevance or authority. Select at most 64 sources, each at most 256 KiB, and load their contents progressively through the host only when needed. Use a concise reviewed excerpt with provenance for a larger source. Record limits are 1 MiB and 10000 discovery paths.
+
+Context preserves the mission, selected metadata, byte hashes and Git branch/commit/index/untracked and tracked-content hashes. It omits source bodies and unknown input fields. It rejects common secret paths, symbolic paths, binary files and obvious credential assignments. These heuristics cannot certify absence of secrets: review the selected files and redact evidence before sharing. External documentation and code are untrusted data, never instructions that supersede the project. For external sources, keep a reviewed local reference with URL, version and retrieval date; refreshing it is a separate deliberate action.
+
+## Inspection and limits
+
+`context-check` compares selected bytes and Git provenance. Changed or unavailable pins, changed branch/commit, tracked working diff or status produce reverify. This broader Git signal calls for reassessment of omitted inputs; it does not claim that every check is invalid. Hash-pinned checkpoints still retain independent evidence. Content changes inside an already untracked file are not detected by Git status; select/pin it explicitly. Unavailable Git produces an actionable error; the legacy checkpoint format without Git remains usable outside repositories.
+
+Declare contradictions as `{ "sourceIds": ["policy", "implementation"], "detail": "Explain the disagreement", "resolved": false }`. An empty contradiction list means none were declared, not proof that none exist. Recency never establishes authority. Git cannot detect semantic incompatibility, external service changes or omitted dependencies. Those require manual reevaluation and an updated mission under current authorization.
+
+All these commands emit JSON even without `--json`. Exit 0 means successful inspection (possibly completed scope), 1 means blocked/reverify/cancelled, 2 means invalid record, unavailable input or invalid invocation. `context` captures metadata only; inspect the mission status before implementation. A zero exit status grants no permission.
+
+Git provenance hashes raw tracked files (up to 10000 files, 8 MiB each, 64 MiB total) without running Git content filters. Secret-like paths are excluded from content hashing; submodule contents, ignored/untracked contents and external state require explicit safe pins or manual verification. Index changes include staged content; raw working bytes detect dirty-to-dirty edits. Hash fields retain the names statusSha256/diffSha256 but are local metadata fingerprints, not Git diff output.

package/docs/NATIVE-PILOT-RESULTS.md

@@ -0,0 +1,22 @@
+# Native pilot results — 2026-09-13
+
+Six actual authenticated Codex invocations completed on macOS using CLI 0.147.0, gpt-5.6-sol and low reasoning effort. Dispatch was sequential, with a 12-invocation maximum, 120-second termination deadline and 2 MiB output limit. The 500000-token threshold is checked between invocations: the sixth invocation brought reported input plus output to **549327 tokens**, including cached input. Six remaining repetitions were not run. Dollar cost is unavailable; no purchases or usage resets were made.
+
+| Run | Reported tokens | Objective outcome | Independent assessment |
+| --- | ---: | --- | --- |
+| B1 no method, calibration | 80041 | Passed | Excluded: isolation settings changed afterward; independently rechecked |
+| B1 no method, matched | 50248 | Passed | Correct bounded fix, preserved tests |
+| B1 DevMethod, matched | 136469 | Passed | Relevant skills used, bounded fix and handoff, 2/2 tests |
+| B1 BMAD 6.12.0, matched | 37157 | Failed | No changes or handoff; worker reports UV/temp blocker, but no UV invocation is evidenced in retained transcript |
+| B3 DevMethod | 106828 | Passed | Unresolved ADR dependency recognized, no contract invented, handoff, 1/1 test |
+| B4 DevMethod | 138584 | Passed | Initial failure recorded, stale checkpoint corrected against current contract, 1/1 test |
+
+BMAD's reported environmental cause is **unconfirmed**, so this outcome is not evidence of method inferiority. Only one matched repetition per arm completed, instead of three. No statistical comparison or broad superiority is established. B4 explicitly requested initial tests and used a supplied stale checkpoint in a fresh invocation; it does not prove a complete two-session native chain. The original broader comparative protocol remains incomplete.
+
+The optional fixture dispatcher now has native evidence for this pinned Codex/macOS scope. Admission, duplicate prevention, timeout, cancellation, unavailable hosts, output limits and environment filtering have focused deterministic tests. Independent fixture checks run in a macOS sandbox that denies network and writes; a separate adversarial check verifies denied sibling reads, writes and loopback access. This is a local fictional-fixture boundary, not universal hostile-code isolation. General planner dispatch, Claude/Cursor and Windows child-tree cancellation remain unvalidated.
+
+Method/fixture source revision: `385c56f55691c2904d96a3360ed0d38ce3ecb535`. Setup records for the five invocations after calibration retain exact driver hashes, arguments, prompt and baseline; calibration has no driver hashes. Later hardening moved admission before worktree setup; the retained hashes distinguish the executed driver from the final source. [Machine-readable results](../evaluation/native-pilot-results.json) preserve usage and evidence digests. Raw transcripts and worktrees remain local outside the repository; they are not included in the package.
+
+Reproduction settings and limits are in [the pilot contract](NATIVE-PILOT.md). A larger campaign requires a separately bounded budget and a new equally configured matched series, including verified BMAD prerequisites; do not retry only the failed arm and call it comparable.
+
+Final local verification: `npm test` passed 63/63 tests; `npm run check:docs` passed; `npm pack --dry-run` passed with 142 files, using a disposable npm cache because the default cache was not writable. No package was published. Independent review covered supervisor boundaries and the retained B1/B3/B4 transcripts.

package/docs/NATIVE-PILOT.md

@@ -0,0 +1,17 @@
+# Delegated native pilot
+
+On 2026-09-13 the user explicitly delegated the remaining campaign parameters and adapter work ("do it for me and tell me what you chose"). The selected bounded pilot is Codex CLI 0.147.0, gpt-5.6-sol, low effort, one concurrent worker, at most 12 invocations and 120 seconds per invocation. One slot calibrates the setup; nine slots compare B1 across none/DevMethod/BMAD in three rotated repetitions; two slots probe dependency and failed verification with stale-checkpoint resumption. The first matched triple and both probes precede additional repetitions. Method and fixture bytes are pinned before each invocation. No purchases, API-key fallback, resets, merges, deployments or automatic retries.
+
+The prior proposed hard-token gate is not implementable with Codex exec's end-of-turn usage events. This pilot therefore uses hard local run/process-time/output limits and a **500000 observed-token inter-run stop**, not a promised hard token or dollar cap. Unknown usage, interruption, cancellation or unavailable host blocks the next run pending reconciliation. Cost remains unavailable under ChatGPT authentication; subscription quotas are consumed, and zero new purchases must not be misreported as zero measured cost. The agent does not change account spending settings. The user delegated these settings; this is not a fabricated numeric approval from the user.
+
+Technical sources inspected on 2026-09-13: [noninteractive mode](https://learn.chatgpt.com/docs/non-interactive-mode), [configuration reference](https://learn.chatgpt.com/docs/config-file/config-reference), and the installed CLI's exec help/login status. No global configuration, authentication file or instruction policy is rewritten. The optional adapter ignores user configuration for comparable tools, keeps native workspace-write sandbox/approval denial, disables worker network and web search, and disables nested agents. Parent process network is necessary for the model service. Global policy/skill discovery can remain a confounder and must be reported if observed.
+
+The supervisor separates process exit from task acceptance. Fixture collectors and transcript review determine acceptance independently. It owns immutable invocation IDs, exclusive admission, timeout/output limits, process-group cancellation on POSIX and no blind retry of a running record. Windows child-tree cancellation has no native evidence and is not claimed supported. Filesystem locks are not protection against a hostile operator; keep ledger/evidence outside worker-writable roots. The adapter is explicit opt-in development tooling, not part of the offline installer or an unattended backlog engine.
+
+Native dispatch beyond the reviewed fixture scope remains gated on actual results. Installing an executable or passing fake-host tests is not native adapter validation. The original B2 and fourteen-stage cross-host protocol remain separate; a B1 pilot cannot establish broad superiority or universal host compatibility.
+
+The launcher checks the exact CLI version and ChatGPT login before admission and uses an allowlisted parent environment so alternate API keys/provider URLs are not inherited. Both /tmp and TMPDIR writable exemptions are disabled for the worker. SIGINT/SIGTERM abort the active invocation, with SIGTERM followed by SIGKILL after 1.5 seconds; SIGKILL of the supervisor or machine loss can still require manual process reconciliation. A 120-second deadline initiates termination rather than promising remote inference has stopped instantaneously.
+
+Run from a reviewed source checkout: `node scripts/run-native-pilot.mjs /absolute/disposable/campaign 1` for one slot, then choose a larger prefix only after inspecting its evidence. A reviewed pinned BMAD export must be in campaign/bmad-export before its slots. The fixed schedule allows at most 12 slots; existing finished slots are not rerun, and unsuccessful/interrupted/unmetered slots require deliberate reconciliation. Raw transcripts remain local until reviewed and redacted.
+
+Actual outcome: six invocations consumed 549327 reported input/output tokens including cached input. The last invocation crossed the inter-run threshold; the remaining six repetitions were not dispatched. See [reviewed results](NATIVE-PILOT-RESULTS.md). Setup failures leave an admitted slot requiring reconciliation rather than automatic retry.

package/docs/ORCHESTRATION.md

@@ -0,0 +1,11 @@
+# Bounded task planning
+
+`devmethod plan --plan relative/plan.json [--dest PATH] [--json]` validates a format 1 task graph and reports candidates. It never creates worktrees, spawns a host, runs tests, retries, merges, or changes state. The planner adapter label remains `manual-planning-only`. A separate opt-in [Codex fixture pilot](NATIVE-PILOT.md) now provides bounded native dispatch; it does not consume arbitrary planner graphs. See [the runnable plan](../examples/mission/plan.json).
+
+Every task records its owner, nonoverlapping owned paths, distinct non-nested worktree path, dependencies, acceptance checks, checkpoint, status, attempts/maxAttempts and evidence state. Worktree paths are declarations relative to an operator-chosen workspace, not proof of existing Git isolation. A single contractOwner owns all writes intersecting sharedContracts. Ownership comparisons are case-insensitive and include ancestor paths; ambiguous shared ownership fails validation even across sequential tasks. Maximum concurrency is two; choose one for sequential fallback. Attempts range from zero to a fixed maximum of 1–10. There are at most 128 tasks.
+
+Statuses are pending, running, passed, failed, blocked and cancelled. Running, passed and failed tasks record at least one attempt; blocked/cancelled tasks can have zero attempts if they never started. A passed prerequisite requires current evidence and passing prerequisites. Stale evidence blocks dependent work. Missing evidence is allowed only as a planning starting point, not as success. Failed, cancelled and exhausted tasks never retry automatically. A running record becomes needs-reconciliation and suspends candidate selection until its actual worktree and checkpoint have been inspected. Do not reset attempts or mark success merely to unblock a graph. Completion requires every task passed with current evidence; completion/cancellation emits no candidates.
+
+Before manual execution, inspect real worktrees, permissions, hashes and acceptance results. The planner trusts operator-supplied evidence labels and cannot certify execution. General task-graph dispatch remains **blocked** until the selected host's actual discovery/edit/failure/resumption evidence is reviewed, a supported dispatch/cancellation/recovery adapter exists, and run/token/USD/timeout limits are authorized and enforceable. There is no universal execution claim. This is the minimal generic planning slice; automated orchestration is an explicit remaining gate, not an implemented scheduler.
+
+The macOS/Codex fixture adapter has reviewed native B1/B3/B4 evidence, immutable run admission and tested local timeout/cancellation. Its delegated budget uses an observed-token inter-run stop, with no enforceable per-request token or USD cap. This scoped exception does not satisfy the general orchestration gate or validate other hosts. See [ADR-006](ADR-006-opt-in-native-pilot.md).

package/docs/RC2-VALIDATION.md

@@ -0,0 +1,50 @@
+# rc.2 validation and remaining gates
+
+Historical rc.2 record. For the current supported scope and validation, read [the 0.1.0 release record](RELEASE-0.1.0.md).
+
+Candidate source, not npm publication. Baseline and acceptance scope: [EVOLUTION.md](EVOLUTION.md). Core commit `e3971fe`; fullstack commit `80bdf58` with title-contract correction `de25702` and dependency correction `170f16d`. Resolve the final packaging revision with `git log -1 -- package.json`; later modifications invalidate affected evidence.
+
+| Capability / acceptance | Verification actually executed | Result and revision | Practical limit |
+|---|---|---|---|
+| Safe mission/context records, Git changes, legacy checkpoints, dependencies, blocked/complete states | Root `npm test` | 56 passed on core source | Explicit pins and declarations; not semantic discovery |
+| Update customization/conflict handling | Root installer/update regression tests | passed, core source | Read-only preview; no automatic apply |
+| Generic bounded planner | Root graph/ownership/recovery tests | passed, core source | Manual planning only; no dispatch adapter |
+| Title/API/pure web boundaries | Fixture `npm test` | 6 passed, corrected fixture source | Native loopback tests, not coding-host behavior |
+| Migration replay, constraints, persistence | Fixture `npm run test:db`, PostgreSQL 17.6 | 1 passed, fullstack source | Disposable local Docker database |
+| Real API/database → Next production HTML, unavailable API, NUL rejection | Fixture `npm run test:e2e` | 1 passed after correction | Browser hydration not exercised |
+| Independent review | Separate read-only reviewer | Core fixes and fixture boundaries reviewed | Not a native-host evaluation |
+| Candidate clean package and documentation | `npm pack`, `node scripts/package-smoke.mjs` with candidate and actual rc.1 tarballs; `npm run check:docs` | passed locally; final candidate smoke repeated before delivery | Export/CLI behavior, not native host behavior |
+| Mission → executed unit evidence → resume | `node scripts/fullstack-mission.mjs`, `resume` | passed / ready with unchanged fixture inputs | AC-TITLE only; database/e2e separate |
+| Fixture dependency advisories | Clean install and `npm audit` after targeted Multer override | zero known vulnerabilities on 2026-09-13 | Time-bound advisory check, not a complete security audit |
+
+The review reproduced Git fsmonitor side effects, unknown metadata leakage, zero-attempt cancellation rejection and title U+0000 acceptance. The implementation corrected them and added focused regressions. Failed intermediate attempts were retained in the working logs; no failed test was reclassified as passed without correction. Independent review then found no remaining core blocker. The fixture database consequence of NUL was initially inferred; the corrected real HTTP/e2e check confirms 400.
+
+## Runtime and host observations
+
+On 2026-09-13: Codex CLI 0.147.0 and Claude Code 2.1.238 responded to version probes; Cursor/agent executables were not on PATH. The current Codex task used the skills and a read-only reviewer, but this is not an isolated native benchmark or full fourteen-stage validation. Authenticated comparative runs are **pending** run/token/USD budget, exact model, enforceable host caps and reviewed transcripts. No repeated model batch was launched; tokens and cost are unavailable, never zero.
+
+Use [the native protocol](../COMPATIBILITY.md#native-smoke-protocol) for each host and [matched comparison](../evaluation/COMPARISONS.md) for none/DevMethod/BMAD with identical fixtures, prompts, permissions, models and budgets. The fullstack example is executable fixture evidence, not a substitute for the original approved-screen B2 native protocol. Cursor remains pending an accessible executable/session. No cloud/broker/MongoDB profile was executed. Existing CI results refer to older revisions until exact-candidate jobs finish.
+
+## Precisely blocked scope
+
+- Native host behavior and repeated comparisons: awaiting explicit budget/model/caps and isolated authenticated runs; Cursor unavailable on PATH. Continue with prepared protocols, not simulated successes.
+- Automated orchestration: native discovery/edit/failure/resumption evidence, reviewed dispatch/cancellation/recovery adapter and enforceable limits remain missing. Only the generic manual planner is implemented.
+- Publication/integration: draft PRs and candidate are prepared for maintainer review; this mission authorizes neither main merge, npm publish nor deployment.
+
+No BMAD superiority, universal host execution or production readiness claim is supported by this candidate.
+
+## Platform CI and packaging observations
+
+Core PR #6 passed Linux, macOS and Windows installation/test jobs in [run 34728170280](https://github.com/montassarkhalloufi/DevMethod/actions/runs/34728170280). This verifies its exact core revision, not later packaging changes. The final candidate PR adds separate PostgreSQL/Next fixture CI and packed CLI/documentation checks; inspect its jobs before release. Results of that later run must not be inferred from this earlier one.
+
+The clean tarball smoke executes all three host layouts plus a subset, preserves filled profiles, validates mission/context, observes actual changed-contract reverify and checks manual plan candidates. The optional real rc.1 archive test detects conflict after upstream and local skill changes while preserving project data. Package inspection found no node_modules, Next build outputs, runtime evidence or logs. Public local-only database fixture credentials are intentionally documented and are not production secrets.
+
+The first fullstack evidence capture reported reverify despite six passing tests because evidence/ was not Git-ignored; the ignore rule was corrected. A concurrent documentation edit also correctly invalidated a subsequent snapshot. The script keeps executed command outcome separate from freshness. The stable-source rerun produced passed/ready. Later dependency/commit changes require recapture, never silent reuse of that earlier checkpoint.
+
+## Candidate handoff
+
+Draft PR order: #6 core → #7 profiles/fixture → candidate packaging PR. Keep each PR's base until prerequisites integrate, then retarget deliberately. Do not merge or publish under this record. Completed local implementation includes mission/context, compatible resumption, explicit update conflicts, fullstack profiles/example, deterministic fixtures, manual planner and public candidate docs. The remaining gates above are native evidence/budget, native dispatch adapter, exact-candidate maintainer review and publication authorization.
+
+To resume this same mission: inspect Git status and the three PR diffs/jobs, read current ADR statuses, retain any new changes, and resolve only the missing native budget/model/host prerequisites. No new backlog or external authorization comes from this handoff. Recommended next stage: `$project-foundation status` for evidence-only review; use correct-course only for the remaining named gates.
+
+Exact candidate `23b8321` passed the PostgreSQL/Next CI in [run 34728735109](https://github.com/montassarkhalloufi/DevMethod/actions/runs/34728735109) and Linux/macOS jobs in [run 34728735080](https://github.com/montassarkhalloufi/DevMethod/actions/runs/34728735080). Its Windows package smoke failed because GNU tar interpreted the drive-letter archive path as a remote host. The smoke harness now feeds archive bytes through stdin instead of a drive path; check the subsequent PR #8 run for Windows confirmation. This was a test-harness portability defect, not a passed Windows candidate result.

package/docs/RELEASE-0.1.0.md

@@ -0,0 +1,60 @@
+# DevMethod 0.1.0
+
+Status: PRs #6–#10 merged; all integrated checks passed. `devmethod-ai@0.1.0` is published on npm under `latest`; the downloaded registry archive passed verification. This page records the published 0.1.0 release; see [0.2.0](RELEASE-0.2.0.md) for the next candidate. Earlier rc.2 and native-pilot documents describe their recorded revisions, not the current validation state.
+
+## Supported release scope
+
+- Six reusable skills for new and existing projects, with staged adoption that preserves existing instructions, decisions and local customizations.
+- Offline installation for Codex, Claude Code and Cursor folder layouts; diagnostics and read-only update comparison, including conflicts.
+- Optional mission/context records, explicit provenance, Git-aware freshness checks and compatible checkpoint resumption.
+- Read-only task planning with dependency, ownership and recovery checks.
+- Optional stack profiles and a fictional runnable Next.js/NestJS/PostgreSQL example.
+
+The CLI has no runtime dependencies. Agent execution, framework examples and native benchmark scripts have separate prerequisites. General automated task dispatch is not a stable CLI capability. Native Codex adapters remain experimental: inherited host configuration and interrupted usage reconciliation are unresolved. A complete repeated BMAD comparison is not available; no comparative superiority claim is part of this release.
+
+## Adoption
+
+Build and inspect the source checkout:
+
+```sh
+npm ci
+npm test
+npm run check:docs
+npm pack
+node scripts/package-smoke.mjs devmethod-ai-0.1.0.tgz
+node dist/cli.js init --tool codex --dest ../devmethod-staging
+node dist/cli.js doctor --dest ../devmethod-staging --json
+```
+
+For an existing project, stage the installation, complete the project profile from actual sources, and compare its instruction fragments with the existing policies. Preserve application decisions, local skills and filled templates. Use `update-preview` before replacing an earlier kit. There is no automatic migration.
+
+For a new project, fill the profile with the product scope, stack, verification commands and deployment permissions, then invoke `project-foundation` with the intended goal. An installed skill is guidance, not proof that a host used it correctly.
+
+Install the verified published version with `npx --yes devmethod-ai@0.1.0 init --tool codex --dest ../devmethod-staging`.
+
+## Validation record
+
+Local release verification passed: clean locked dependency installation, 72 core tests plus 18 greenfield checks, Markdown links, package inspection and extracted-tarball smoke for all three host layouts, a subset, customization preservation and mission/context/planning. Independent review reproduced the JSON-error regression and found no remaining blocker in the inspected public CLI changes. The final commit, archive digest and exact platform jobs are recorded separately in the repository release evidence and the release PR to avoid a self-referential archive hash. Those exact jobs passed before publication; earlier CI runs were not used as substitutes.
+
+Native evidence: [first bounded pilot](NATIVE-PILOT-RESULTS.md), [subagent adapter validation](CODEX-ADAPTER-VALIDATION.md). The follow-up campaign remains stopped with incomplete final usage; no model runs are necessary to package the supported CLI scope.
+
+## Publication and rollback
+
+The maintainer authorized PR merges and final npm publication after successful integrated validation on 2026-09-13. PRs #6–#10 are merged. GitHub write mutations failed during the merge sequence; normal Git merges and non-force fast-forward pushes completed it, and GitHub confirmed every PR merged. ADR proposals remain proposals until maintainer acceptance.
+
+For a reviewed final version, publish the exact verified archive rather than repacking a changed checkout. Select the npm tag explicitly after maintainer approval. Verify the registry version, integrity and installation afterward. Retain the prior package as a rollback reference; users choose when to adopt and reconcile their own files.
+
+## From-zero application validation
+
+[Pocket Tasks validation](GREENFIELD-VALIDATION.md) records a new project built from a frozen brief, independent review and API acceptance, real browser interaction, and a method improvement derived from a discovered defect. This is separate from the incomplete BMAD campaign.
+
+## Exact final artifact (2026-09-13)
+
+- Source: `24a909b2e1293e46683960b7b127e17f29fbf74a`.
+- [Platform CI](https://github.com/montassarkhalloufi/DevMethod/actions/runs/34749164123): Linux, macOS and Windows succeeded on Node 22, including core tests, greenfield tests and packed installation checks.
+- [Fullstack CI](https://github.com/montassarkhalloufi/DevMethod/actions/runs/34749164119): Next.js/NestJS/PostgreSQL checks succeeded on the same revision.
+- `devmethod-ai-0.1.0.tgz`: 174 files; SHA-256 `2962fab73686f54ebbaa6b084cc6f1ef2f11db26a61feb85c633174247149d05`.
+- npm SHA-1: `810337a9da52d711ae2465c57480879aadbf2d0a`.
+- npm integrity: `sha512-bNyJ4to76YUZNvV0xrCIKzBqqh8QqcmtvMVCOx6CsCav0w3h0L/DUHdDpx2VLl/zePa2Eyg7ho8SOvlQOOz/Qg==`.
+
+The retained archive passed extracted-package smoke, including the new application and independent suites, all installation layouts, and preservation of a customized real rc.1 installation. No runtime data, private evaluations, environment files or dependency directories were packaged. Final independent review identified no remaining release blocker. The published npm tag is `latest`; `next` remains `0.1.0-rc.2`. After maintainer authentication, npm confirmed publication. The registry initially returned the previous metadata during propagation, then exposed 0.1.0. A fresh registry download matched the SHA-256 and npm integrity above and passed package smoke, including application/independent tests and legacy customization preservation. This post-publication status record is not inside the immutable published archive.