v0.0.14 — v0.0.13 recovery + scaffold/CLI ergonomics

[0.0.14] — 2026-05-08

Theme: v0.0.13 recovery — 3 must-fix correctness bugs + 6 scaffold/CLI ergonomics polish. v0.0.13's URL-shortener BASELINE surfaced 3 hard regressions (apostrophe-strip phantom no-converge, workspace:* shipped to npm, factory spec review broken on published packages) plus 6 honorable-mention polishes. v0.0.14 ships all 9 in 8 sibling specs — 7/8 first-iter, 1/8 agent-exit-nonzero with work landed (same v0.0.11 worktree-spec friction shape; the implementation is correct).

Added

• Harness regex-no-match safety net (factory-harness). When bun test -t <pattern> exits nonzero AND its output contains regex .+ matched 0 tests, the runner classifies the satisfaction as status: 'error' with detail prefix harness/test-name-regex-no-match: <pattern> matched 0 tests in <file> instead of 'fail'. The runtime treats error as a tooling-mismatch signal (not as "implementation broken — re-implement"), which prevents the v0.0.13 phantom-no-converge loop. Mirrors v0.0.13's coverage-trip detection shape.
• factory finish-task --all-converged status-aggregator (factory-core). Walks each candidate factory-run's descendant factory-phase records, finds the FINAL iteration's terminal phase, and only ships when terminalPhase === 'pass'. Skipped specs surface as factory: skipped <id> (run <runId> did not converge — last phase: <status>). The all-no-converge sequence case is a safe no-op (exit 0, shipped=0). Closes the v0.0.13 BASELINE friction where finish-task disagreed with run-sequence on what "converged" means.
• resolveContextDir helper (factory-core). Pure function exported from @wifo/factory-core that resolves the precedence chain cliFlag → configValue → './.factory'. Used by factory finish-task; mirrored in factory-context. Public API surface delta: @wifo/factory-core 34 → 35 (+1).
• Day-zero scaffold passes its own DoD gates (factory-core). factory init now writes a stub src/index.ts (export const VERSION = '0.0.0';) + src/index.test.ts (1 test) so pnpm typecheck && pnpm test && pnpm check all exit 0 on the freshly-scaffolded tree. Single-line BIOME_JSON_TEMPLATE arrays so biome's lineWidth=100 rule doesn't trip on the template's own format. The agent overwrites the stubs on first feature scope.
• Lint code spec/yaml-colon-needs-quoting (factory-core). Severity warning. Detects unquoted frontmatter values containing : (which YAML parses as a nested mapping) and emits a friendly fix-suggestion ("wrap the value in single quotes"). /scope-project's authoring rules updated to auto-quote — prevents the bomb at scoping time so the lint warning rarely fires in practice.
• --setting-sources project,local on claude -p spawns (factory-runtime). Excludes user-level ~/.claude/settings.json global plugin/skill auto-suggestion hooks (e.g., the Vercel/Next.js skill injection the v0.0.13 implement-phase agent reported as false-positive noise). Subscription auth is preserved (OAuth/keychain credentials live outside settings.json). Project-level .claude/settings.json and *.local.json hooks still load — only user-level global noise is filtered.

Changed

• normalizeTestNamePattern no longer strips apostrophes (factory-harness). v0.0.12's apostrophe-stripping (added to handle hypothetical agent stylization of it() names) caused a real soundness bug: specs that genuinely used "slug's log" had the apostrophe stripped to "slugs log" before passing to bun's -t regex, but the actual test still had the apostrophe → bun matched 0 tests → 5 wasted iterations on phantom no-converge in the v0.0.13 BASELINE. The strip set is reduced to curly-double-quotes only; apostrophes (ASCII + curly) are now treated as literals. Modern Claude preserves apostrophes correctly in test names; the original motivation didn't manifest.
• factory spec review dispatches via child_process.spawn instead of createRequire (factory-core). v0.0.13's createRequire(import.meta.url) workaround for the workspace cycle hit CJS resolution against factory-spec-review's ESM-only exports map → silent exit 0 on the npm-published artifact (the workspace symlink-bypass masked it). v0.0.14 spawns the factory-spec-review bin as a subprocess; process boundary eliminates the type-resolution AND the CJS/ESM mismatch. Subscription auth preserved via env propagation. The factory-cli umbrella package option stays deferred for v1.0.0.
• --context-dir universal default flips from ./context to ./.factory (factory-core, factory-context). factory init creates ./.factory, so the previous split — factory-runtime defaulted to ./.factory while factory finish-task and factory-context defaulted to ./context — silently looked in the wrong directory. v0.0.14 unifies all three CLIs on ./.factory AND adds auto-detect from factory.config.json runtime.contextDir. Precedence: CLI flag (--context-dir) > factory.config.json runtime.contextDir > universal default ./.factory. Breaking for users who relied on the old ./context default — pass --context-dir ./context to keep the v0.0.13 behavior.
• CI publish flow uses pnpm pack + npm publish <tarball> two-step (workflow). v0.0.13.x's switch from pnpm publish to npm publish (for OIDC handshake) lost pnpm's automatic workspace:* → ^<version> rewrite, breaking npx @wifo/factory-core init for fresh users. v0.0.14 splits the responsibilities: pnpm pack does the manifest rewrite + produces a tarball; npm publish <tarball> --provenance uploads via OIDC. Post-publish verification re-fetches each manifest and fails the workflow if workspace: leaks through — regression-pin against future shape regressions.
• All six @wifo/factory-* packages bumped to 0.0.14 in lockstep. init-templates.ts PACKAGE_JSON_TEMPLATE.dependencies bumped from ^0.0.13 to ^0.0.14 for every @wifo/factory-* dep.

Public API surface

Package	v0.0.13	v0.0.14
@wifo/factory-core	34	35 (+resolveContextDir)
@wifo/factory-context	18	18
@wifo/factory-harness	~16	~16
@wifo/factory-runtime	26	26
@wifo/factory-spec-review	10	10
@wifo/factory-twin	~7	~7

Reconciliations worth knowing

• 8 specs across 2 run-sequence invocations. Run-1 hit the same stale-dist issue as v0.0.12 (the runtime process loaded harness/dist at startup with the OLD apostrophe-strip code; spec A's fix landed mid-run but the running process never reloaded). Spec D phantom-no-converged on its own apostrophe-bearing test name; sequence stopped. Recovery: rebuild dist, fresh runtime process, re-run on remaining 5 specs (4/5 converged; spec H — claude-no-hooks — agent-exit-nonzero with work landed, same v0.0.11 worktree-sandbox friction shape). All 8 specs' work is correct; 232 core / 71 harness / 255 runtime / 106 spec-review tests pass.
• --context-dir change is breaking. Users with scripts hard-coded to ./context need to pass --context-dir ./context explicitly, OR rename their dir to ./.factory. The universal default is the dir factory init creates.
• createRequire dropped from core/cli.ts. v0.0.12's workaround → v0.0.13's persistence → v0.0.14's removal. The subprocess path eliminates the CJS/ESM friction entirely.
• bun stays as the test runner. Every package's test script is bun test src — unchanged. v0.0.13's bun-as-test-only documentation continues to apply.
• v0.0.14 explicitly does NOT ship: the factory-cli umbrella package (deferred to v1.0.0); custom hook authoring (out of scope; --setting-sources is enough); a --legacy-context-dir-default flag (rejected — the migration is one flag added per invocation); claude --no-hooks (the actual flag that exists is --setting-sources; we use it).

---

v0.0.13 — init-script ergonomics + brownfield polish + cycle-break

[0.0.13] — 2026-05-06

Theme: init-script ergonomics + brownfield-adopter polish + architectural cycle-break. v0.0.13 closes 8 BACKLOG entries surfaced across the v0.0.12 BASELINE (5 init-script frictions + 1 carve-out re-shape) and the v0.0.12 tag-fire (2 architectural items). Six sibling specs ship together — first cluster to converge 6/6 first-iter in a single run-sequence invocation with zero recovery.

Added

• factory init first-contact polish (factory-core). Three init-template fixes that close every greenfield friction the v0.0.12 BASELINE flagged: (a) BIOME_JSON_TEMPLATE migrates from Biome 1.x's "include" key to Biome 2.x's "includes" (matching the pinned major); (b) scaffold writes .factory/.gitkeep + appends .factory/worktrees/, .factory/twin-recordings/ to .gitignore (idempotent); (c) factory.config.json gains dod.template?: string[] derived from package.json scripts at init time, and /scope-project reads it as the canonical DoD section body — closes the per-cluster 4× spec/dod-needs-explicit-command warning the v0.0.12 BASELINE measured.
• Auto-quiet for non-TTY stderr (factory-runtime). When process.stderr.isTTY is false (pipe / tee / CI capture), the [runtime] progress lines auto-suppress. New --no-quiet (and --progress alias) opts back into progress for non-TTY contexts that want it. Precedence: CLI flag > factory.config.json runtime.quiet > auto-detect > built-in default. Closes the v0.0.12 BASELINE honorable mention "live progress pollutes captured logs."
• factory finish-task --all-converged (factory-core). Batch ship-cycle move-to-done. Walks the most recent factory-sequence (or one named via --since <factorySequenceId>); moves each converged spec from <dir>/<id>.md to <dir>/done/<id>.md; emits one factory-spec-shipped record per moved spec. Mutually exclusive with the positional <spec-id> form. Closes the only step that didn't auto-progress with the rest of run-sequence.
• Per-scenario coverage-trip detection (factory-harness). v0.0.12's option (a) (--coverage=false) was descoped because bun 1.3.x rejects the flag; v0.0.13 ships option (b) — the runner parses bun's stdout for 0 fail + coverage threshold ... not met markers AND a nonzero exit, classifying the result as pass with detail prefix harness/coverage-threshold-tripped: rather than fail. Real test failures still classify as fail; the carve-out is conservative.
• Bun-as-test-only — schema emitter is Node-native (factory-core). packages/core/scripts/emit-json-schema.ts rewritten to use node:fs + standard ESM. Build script changes from bun run scripts/... to tsx scripts/.... tsx joins devDependencies of @wifo/factory-core. pnpm build and pnpm typecheck are now Node-only — bun is required only for pnpm test (the test runner is bun test src per package — intentional and unchanged). Documented in top-level README + AGENTS.md + scaffold README.

Changed

• @wifo/factory-spec-review is now a peerDependency of @wifo/factory-core (was dependencies in v0.0.12). pnpm 8+ and npm 7+ auto-install peer deps, so the v0.0.12 brownfield zero-config goal is preserved for the major package managers. The build-graph cycle introduced in v0.0.12 disappears: peer deps don't form build-time edges. Legacy npm < 7 caveat documented in packages/core/README.md + top-level README.
• packages/core/src/cli.ts reverts to a static import { runReviewCli } from '@wifo/factory-spec-review/cli' — the v0.0.12 createRequire(import.meta.url) workaround is removed (the cycle is gone, no need to defer type resolution).
• .github/workflows/publish.yml build step reverts to pnpm -r build (single line). The v0.0.12 explicit per-package list and --workspace-concurrency=1 flag are removed. Build-before-typecheck order preserved (independent constraint).
• All six @wifo/factory-* packages bumped to 0.0.13 in lockstep. init-templates.ts PACKAGE_JSON_TEMPLATE.dependencies bumped from ^0.0.12 to ^0.0.13 for every @wifo/factory-* dep.

Public API surface

Package	v0.0.12	v0.0.13
@wifo/factory-core	34	34
@wifo/factory-context	18	18
@wifo/factory-harness	~16	~16
@wifo/factory-runtime	26	26
@wifo/factory-spec-review	10	10
@wifo/factory-twin	~7	~7

Reconciliations worth knowing

• v0.0.13 is the first cluster to ship 6/6 first-iter via one run-sequence invocation with zero recovery. Total: 71m wall-clock, 169k charged tokens. v0.0.10 (5/5 in 81m) was the prior best; v0.0.11 had iter-2 retry on one spec; v0.0.12 took 4 invocations due to --coverage=false recovery. The v0.0.13 ship is the cleanest dogfood evidence to date that the runtime is stable.
• Peer-dep auto-install requires pnpm 8+ or npm 7+. Legacy npm users get a clear stderr message; the install-docs caveat is one-line in both top-level README and packages/core/README.md.
• bun is still required for pnpm test. Every package's test script is bun test src — that's the workspace's chosen test runner, intentional and unchanged. The setup-bun GitHub Actions step also stays for the test phase.
• factory init --adopt (shipped v0.0.12) interacts cleanly with the v0.0.13 init-template additions. The IGNORE_IF_PRESENT set still skips pre-existing package.json/tsconfig.json/biome.json. The new .factory/.gitkeep and .gitignore extension are append-only — no overwrite.
• v0.0.13 explicitly does NOT ship: vitest/jest equivalents for coverage-trip detection (deferred — bun-only); factory-cli umbrella package for the cycle-break (option b from the v0.0.13 BACKLOG entry — out of scope; peer-dep is the lighter v0.0.13 solution); migrating tests off bun (out of scope; bun test stays); JSON-line streaming progress format; per-channel progress fd; auto-fire factory finish-task on convergence.

---

v0.0.12 — validate-phase reliability + brownfield onramp + observability + DoD trust

[0.0.12] — 2026-05-05

Theme: validate-phase reliability + brownfield-adopter onramp + observability + DoD trust. v0.0.12 closes 12 BACKLOG entries surfaced across the v0.0.11 ship, the OLH CORE-836 dogfood, and the v0.0.11 short-url BASELINE. Six specs ship together: quote normalization + lint warning in the harness; cause-of-iteration / live-progress / tooling-mismatch detection in the runtime; dedup-status correctness + filesChanged debug telemetry + agent-exit stderr-tail capture; literal-command DoD trust contract; factory init --adopt + factory finish-task + factory-spec-review as hard dep; smoke-boot scenarios in /scope-project.

Added

• Harness quote normalization + scoping-time lint (factory-harness, factory-core). parseTestLine strips/replaces curly quotes / smart quotes / backticks before substring match, so a stylistic apostrophe drop between spec and test no longer no-matches correct work. New spec/test-name-quote-chars lint warning catches the friction at scoping time.
• Live progress on stderr (factory-runtime). One stderr line per phase boundary (start + end with timing/charged-tokens/files-changed). Cause-of-iteration line at iter N+1 start: [runtime] iter <N> implement (start) — retrying: <K> failed scenarios; <M> failed dod gates. New --quiet flag (and factory.config.json runtime.quiet) suppresses progress; existing stdout (sequence summary, dynamic-DAG promotion logs) preserved.
• Tooling-mismatch loop detection (factory-runtime). Warning line on iter N+1 when iter N-1 and iter N have dod.status === 'pass' AND identical validate.failedScenarios. Fires once per run; signals "the runtime is iterating against a non-real failure" without changing behavior.
• Dedup correctness (factory-runtime). run-sequence's already-converged dedup now aggregates each candidate factory-run's descendant factory-phase records and verifies every iteration's terminal phase is status: 'pass' before adding to the skip-map. Closes the v0.0.11 ship bug where a NO-CONVERGE run was incorrectly skipped on retry. New stdout log: factory-runtime: <id> prior factory-run found but status=no-converge — re-running.
• filesChanged debug telemetry (factory-runtime). factory-implement-report.payload gains optional filesChangedDebug: { preSnapshot, postSnapshot } so the next reproduction of the v0.0.11 short-url BASELINE undercount is trivially diagnosable from the persisted record.
• Agent stderr tail capture (factory-runtime). factory-implement-report.payload.failureDetail gains optional stderrTail: string (last 10 KB byte-truncated) when claude -p exits non-zero. Closes the v0.0.11 worktree-sandbox-spec investigation gap.
• Literal DoD shell commands (factory-core, factory-runtime). New spec/dod-needs-explicit-command lint warning fires when DoD bullets look like runtime gates without a backtick command. dodPhase drops the script-name-guessing path; bullets without backtick commands are reported as status: 'skipped', reason: 'dod-gate-no-command-found'. SPEC_TEMPLATE.md updated with worked examples (typecheck clean (\pnpm typecheck`)` etc).
• factory init --adopt (factory-core). New mode that walks the templates but skips files that already exist (logged), creating only factory-specific bits (docs/specs/, docs/specs/done/, docs/technical-plans/done/, factory.config.json, .claude/commands/scope-project.md, .gitignore appended if missing). Brownfield-adopter onramp; closes the CORE-836 friction.
• factory finish-task <spec-id> (factory-core). New CLI subcommand + library helper. Moves the named spec from <dir>/<id>.md to <dir>/done/<id>.md and emits a factory-spec-shipped context record. Refuses to move when no converged factory-run exists. Runtime emits a factory-runtime: <id> converged → ship via 'factory finish-task <id>' hint on convergence (read-only — does not auto-mutate the working tree).
• @wifo/factory-spec-review as hard dep of @wifo/factory-core (factory-core). Single install (npm i @wifo/factory-core or npx @wifo/factory-core ...) brings the reviewer. Drops the lazy spec/review-unavailable error path. Closes the CORE-836 resolver-bounce friction in CodeArtifact-pinned repos.
• Smoke-boot scenarios in /scope-project (factory-core/commands). The slash command source gains a ### HTTP entrypoint smoke-boot scenarios subsection: when a spec mentions createServer / listen(<port>) / app.listen / http.createServer / Bun.serve / serve(, the scoper appends a smoke-boot scenario that boots bun src/main.ts, probes a route, and kills the process. Forces the entrypoint into existence — closes the v0.0.11 short-url BASELINE gap where library code shipped but bun src/main.ts 404'd.

Changed

• All six @wifo/factory-* packages bumped to 0.0.12 in lockstep. init-templates.ts PACKAGE_JSON_TEMPLATE.dependencies bumped from ^0.0.11 to ^0.0.12 for every @wifo/factory-* dep.

Public API surface

Package	v0.0.11	v0.0.12
@wifo/factory-core	33	34 (+finishTask)
@wifo/factory-context	18	18
@wifo/factory-harness	~16	~16
@wifo/factory-runtime	26	26 (+RunOptions.quiet?: boolean field)
@wifo/factory-spec-review	10	10
@wifo/factory-twin	~7	~7

Reconciliations worth knowing

• Per-scenario coverage carve-out descoped to v0.0.13. Original BACKLOG option (a) called for --coverage=false on per-scenario bun test invocations. The implement run found bun 1.3.x rejects this flag (The argument '--coverage' does not take a value.); bun has no CLI override for coverage. Re-opens with option (b) — parse 0 fail + nonzero exit as a coverage trip — for v0.0.13.
• DoD non-LLM dispatch when ANTHROPIC_API_KEY is unset. v0.0.10's DoD-verifier dispatches non-shell judge bullets to the SDK and fails when the env var is unset. The v0.0.12 cluster shipped via --skip-dod-phase; pre-flight factory spec lint + factory spec review still gated quality. The literal-command DoD trust contract (this cluster) reduces the LLM dispatch surface so future runs need fewer judge calls.
• factory init --adopt does NOT mutate the host's package.json. Adding factory devDeps is the maintainer's responsibility. A future v0.0.13 candidate adds --write-deps for opt-in.
• factory finish-task is opt-in, not auto. Move-to-done is a ship action; the runtime emits the hint but never auto-mutates the working tree from inside a phase.
• v0.0.12 explicitly does NOT ship: per-scenario coverage carve-out (descoped — see above); auto-resolution on monotonic-DoD-pass loops (warning only — needs evidence on determinism); filesChanged algorithm replacement (telemetry first; v0.0.13 candidate); JSON-line streaming progress format (deferred); factory init --merge / --dry-run modes; factory.config.json dod.commands map.

---

v0.0.11 — trust → isolation: worktree-sandbox + trust-layer calibration

[0.0.11] — 2026-05-04

Theme: trust → isolation — runs sandboxed in their own git worktree, plus calibration of the v0.0.10 trust-contract layer. v0.0.11 closes the long-deferred worktree-sandbox candidate (BACKLOG since the original ROADMAP's v0.1.0 list) and ships five sibling polish specs that pay down friction surfaced by the v0.0.10 BASELINE: holdout-aware convergence, DoD-precision calibration, dynamic DAG walk for run-sequence, charged-token budget surfacing, and CI publish hygiene.

Added

• factory-runtime run --worktree (factory-runtime). Each run materializes an isolated git worktree at <projectRoot>/.factory/worktrees/<runId>/ on a throwaway branch factory-run/<runId>; the implement / validate / DoD phases all execute against that checkout, so the maintainer's main tree is never touched by the agent. Strong undo by construction (git worktree remove); enables parallel runs against distinct worktree paths. New CLI subcommand factory-runtime worktree { list | clean [--all] [--keep-failed] } for forensic inspection + maintenance. New context record type factory-worktree (parents=[runId]) carrying runId / worktreePath / branch / baseSha / baseRef / createdAt / status. Public surface: +3 exports (23 → 26: createWorktree, WorktreeOptions, CreatedWorktree). New RuntimeErrorCode: runtime/worktree-failed (14 → 15).
• Holdout-aware convergence (factory-runtime). validatePhase({ checkHoldouts: true }) runs ## Holdout Scenarios each iteration alongside the visible scenarios; convergence requires both to pass. New CLI flag --check-holdouts and factory.config.json key runtime.checkHoldouts. The persisted factory-validate-report payload gains a holdouts: [] array (legacy records without it still parse).
• DoD-precision calibration (factory-spec-review). The dod-precision judge is recalibrated against v0.0.10 BASELINE evidence; threshold + prompt updated to reduce false-positive flags on prose DoD bullets.
• Charged-token budget surfacing (factory-runtime). RunReport.chargedTokens now exposes the budget-relevant total (input + output) — cache reads/creates are excluded per Anthropic's pricing. CLI converged/no-converge stdout prints charged=<n>/<budget>. The existing RunReport.totalTokens is preserved as a deprecated alias for back-compat.
• Dynamic DAG walk for run-sequence (factory-runtime). factory-runtime run-sequence now promotes drafting specs dynamically as their deps converge during the same invocation (auto-promotion log line: <dep> converged → promoting <dependent>). Single-invocation cluster shipping for ready→drafting chains; --include-drafting (and runtime.includeDrafting) preserve the legacy walk-everything-from-start semantic.
• factory ci publish hardening (factory-core). CI publish workflow + scripted manual-fallback path normalize tarball checks across all six packages.

Changed

• All six @wifo/factory-* packages bumped to 0.0.11 in lockstep. init-templates.ts PACKAGE_JSON_TEMPLATE.dependencies bumped from ^0.0.10 to ^0.0.11 for every @wifo/factory-* dep.

Public API surface

Package	v0.0.10	v0.0.11
@wifo/factory-core	33	33
@wifo/factory-context	18	18
@wifo/factory-harness	~16	~16
@wifo/factory-runtime	23	26 (+createWorktree, +WorktreeOptions, +CreatedWorktree)
@wifo/factory-spec-review	10	10
@wifo/factory-twin	~7	~7

Reconciliations worth knowing

• --worktree is opt-in for v0.0.11. Default false. Becomes opt-out post-v0.1.0 once soaked.
• One worktree per run() invocation. All phases share the worktree for that run. Cleanup is opt-in via factory-runtime worktree clean (default removes only converged worktrees; --all also removes failed ones, preserving forensic value of failed runs).
• Worktree creation is atomic — a partial failure (disk full, permission denied) leaves no orphan record/branch. NO factory-run record persists when createWorktree fails (the run never started).
• v0.0.11 explicitly does NOT ship: parallel agent execution within a single sequence; auto-cleanup of converged worktrees on run end; worktree GC after N days; auto-merge of converged worktrees back into the main branch (the maintainer reviews + merges manually); per-spec worktree options in spec frontmatter (only RunOptions.worktree for now); cross-platform Windows worktree handling (best-effort; tested on macOS/Linux).

---

v0.0.10 — trust contract: DoD-verifier + reviewer judges + workflow polish

[0.0.10] — 2026-05-03

Theme: trust contract — DoD-verifier + reviewer judges = trust on both sides. v0.0.10 closes the v0.0.9 BASELINE friction list with five sibling specs that together turn each spec's Definition of Done into a checked contract and tighten the spec-quality teeth. The DoD verifier now treats unchecked DoD items as a converge-blocking signal; three new reviewer judges (feasibility, api-surface-drift, scope-creep) catch quality regressions before runtime; run-sequence polish + factory spec watch smooth the workflow; and the spec/wide-blast-radius heuristic is recalibrated against empirical data.

Added

• DoD-verifier phase (factory-runtime). New verifier consumes the spec's ## Definition of Done checklist and surfaces unchecked items as a converge-blocking signal. Public surface: +2 exports (21 → 23).
• Three reviewer judges (factory-spec-review): feasibility, api-surface-drift, scope-creep. Three new ReviewCode union members; export count unchanged at 10.
• factory spec watch (factory-core). Watches a directory tree for *.md changes and re-runs factory spec lint (and optionally factory spec review --no-cache) on every save. Public surface: +2 exports (29 → 31).
• spec/wide-blast-radius threshold raise from 8 to 12 + NOQA suppression directive (factory-core). Threshold raised based on v0.0.8 self-build evidence (the spec that actually timed out touched 12 files) + v0.0.9 BASELINE empirical data (18 historical specs warned at threshold 8). New HTML-comment NOQA directive (<!-- NOQA: spec/wide-blast-radius --> / <!-- NOQA: code-a, code-b --> / blanket <!-- NOQA -->) suppresses warnings per-spec; per-code, per-spec scope; never suppresses severity-error codes.
• run-sequence polish (factory-runtime). Workflow ergonomics improvements landed alongside the other v0.0.10 deliverables.

Changed

• All six @wifo/factory-* packages bumped to 0.0.10 in lockstep. init-templates.ts PACKAGE_JSON_TEMPLATE.dependencies bumped from ^0.0.9 to ^0.0.10 for every @wifo/factory-* dep.
• spec/wide-blast-radius warning message updated to reference the new >= 12 threshold.

Public API surface

Package	v0.0.9	v0.0.10
@wifo/factory-core	29	33 (+watch +DoD helpers)
@wifo/factory-context	18	18
@wifo/factory-harness	~16	~16
@wifo/factory-runtime	21	23 (+DoD-verifier)
@wifo/factory-spec-review	10	10 (+3 ReviewCode union members; count unchanged)
@wifo/factory-twin	~7	~7

Reconciliations worth knowing

• NOQA does NOT suppress errors. Severity-error codes (e.g., spec/invalid-depends-on) always fire. NOQA is warnings-only.
• NOQA placement constraint. The directive lives in the spec body (HTML comment), NOT in the YAML frontmatter — the schema is .strict() and would reject the comment as an unknown key.
• NOQA scope is per-spec, not per-line. A single matching comment anywhere in the body suppresses every emission of the named code from lintSpec.
• No retroactive NOQA additions. v0.0.10 raises the threshold but does NOT walk the 18 historical specs to add NOQA — most fall below 12 paths after the raise; the rest can opt in on demand.

---

v0.0.9 — status-aware run-sequence + per-spec timeout + scaffold scripts

[0.0.9] — 2026-05-03

Theme: status-aware run-sequence + per-spec timeout + scaffold scripts + dep-aware internal-consistency. Four small frictions surfaced in the v0.0.8 BASELINE — drafting specs got pulled into run-sequence; wide-blast-radius specs hit the 600s agent-timeout ceiling with no override; the scaffold's package.json shipped empty scripts even as every spec's DoD claimed they were runnable; internal-consistency flagged shared constraints in multi-spec products as unreferenced because it didn't follow depends-on edges. v0.0.9 closes all four.

Added

• agent-timeout-ms frontmatter field + spec/wide-blast-radius lint warning (factory-core). New optional 'agent-timeout-ms': z.number().int().positive().optional() on SpecFrontmatterSchema. New lint code spec/wide-blast-radius (severity: warning) emitted when a spec's Subtasks section names ≥ 8 distinct file paths. Field-level addition; zero new public exports.
• Scaffold scripts: { typecheck, test, check, build } (factory-core). PACKAGE_JSON_TEMPLATE.scripts now ships the four canonical entries matching the monorepo's own conventions, so a fresh factory init project can immediately run the gates its DoD claims.
• run-sequence skips drafting specs by default + --include-drafting flag (factory-runtime). loadSpecs now filters frontmatter.status === 'drafting' unless includeDrafting: true. New CLI flag --include-drafting and factory.config.json key runtime.includeDrafting (default false). CLI flag > config > built-in default.
• Per-spec agent-timeout-ms consumption (factory-runtime). run() now resolves the agent timeout via spec.frontmatter['agent-timeout-ms'] ?? options.maxAgentTimeoutMs ?? DEFAULT_MAX_AGENT_TIMEOUT_MS. Wide-blast-radius specs override the 600s ceiling without bumping the global default for everyone.
• internal-consistency judge follows depends-on edges (factory-spec-review). The judge's buildPrompt consumes the existing JudgePromptCtx.deps (plumbed through in v0.0.7 by cross-doc-consistency) and appends a ## Deps Constraints (referenced via depends-on) section to the artifact, with each dep's ## Constraints / Decisions block sliced via findSection. The criterion gains a sentence instructing the LLM judge to score against the union of this spec's Constraints + every dep's Constraints. Field-level on existing types; zero new exports.

Changed

• All six @wifo/factory-* packages bumped to 0.0.9 in lockstep. init-templates.ts PACKAGE_JSON_TEMPLATE.dependencies bumped from ^0.0.8 to ^0.0.9 for every @wifo/factory-* dep.

Public API surface

Unchanged across every package. Strictly equal to v0.0.8's surface (29 / 21 / 10 / 18 / ~16 / ~7). All v0.0.9 changes are field-level on existing schemas/types/templates + version metadata.

Reconciliations worth knowing

• internal-consistency cache invalidates on first run after upgrade. The criterion text gains one sentence, so ruleSetHash() flips and v0.0.8 cache entries miss. Expected and correct.
• Dep loading remains the CLI's responsibility. v0.0.7 already wired the CLI to auto-load each declared dep from docs/specs/<id>.md or docs/specs/done/<id>.md and thread them through runReview's deps option. v0.0.9 just teaches internal-consistency to opt into the existing data flow.
• internal-consistency still applies on every spec. The judge's applies() is unchanged (returns true unconditionally). The dep-awareness affects scoring, not whether the judge fires. A spec without deps still gets scored, just without dep-context.
• Transitive dep loading is NOT yet shipped. The CLI loads direct deps only. Walking a dep's own depends-on chain is deferred to a future release.
• Drafting-by-default is run-sequence only. factory-runtime run (single-spec) still runs whatever spec you point it at regardless of status. The status filter is purely a sequence-level guard.

---

v0.0.8 — discoverability + baseline reset

[0.0.8] — 2026-05-03

Theme: discoverability + baseline reset. The v0.0.7 BASELINE run shipped a critical finding: v0.0.7's three deliverables (/scope-project, depends-on, run-sequence) were on npm but invisible to a fresh-repo agent — factory init didn't auto-install the slash command into .claude/commands/, the scaffold README didn't mention run-sequence, and the canonical baseline prompt explicitly told the agent those tools didn't exist (the prompt was authored when they were future work). v0.0.8 closes all three gaps so v0.0.7's value is actually exercised by a fresh-repo agent.

Added

• Bundled /scope-project slash command (factory-core). Canonical source moves from docs/commands/scope-project.md to packages/core/commands/scope-project.md so it ships in the npm tarball. packages/core/package.json's files glob extends with commands. New internal helper readScopeProjectCommandTemplate() in init-templates.ts resolves the bundled markdown via import.meta.url + relative path (works in both source and built contexts). NOT exported from core/src/index.ts.
• factory init writes .claude/commands/scope-project.md (factory-core). The scaffold's planFiles now drops the bundled slash-command source into the fresh project's .claude/commands/ so any Claude Code session opened in the project picks it up zero-config. The file is a regular file (not a symlink — symlinks don't survive npm pack reliably across platforms).
• Scaffold README ## Multi-spec products section (factory-core). New section in init-templates.ts's README_TEMPLATE documents the canonical v0.0.7+ flow: /scope-project <description> → factory spec lint docs/specs/ → factory-runtime run-sequence docs/specs/. Section explicitly notes that factory init writes .claude/commands/scope-project.md automatically. ~25 lines, scannable not tutorial. The scaffold is the documentation.
• Baseline prompt reset (docs/baselines). Archived docs/baselines/url-shortener-prompt.md as url-shortener-prompt-v0.0.5-v0.0.7.md (preserves the historic v0.0.5–v0.0.7 baseline against which prior friction lists were measured). Wrote a fresh canonical prompt opening with /scope-project + factory-runtime run-sequence so v0.0.8+ baselines measure the v0.0.7+ flow honestly. BASELINE.md documents the methodology reset event.

Changed

• All six @wifo/factory-* packages bumped to 0.0.8 in lockstep. init-templates.ts PACKAGE_JSON_TEMPLATE.dependencies bumped from ^0.0.7 to ^0.0.8 for every @wifo/factory-* dep.

Public API surface

Unchanged across every package. Strictly equal to v0.0.7's surface (29 / 21 / 10 / 18 / ~16 / ~7). All v0.0.8 changes are field-level on existing constants/templates + version metadata + new internal-only helpers.

Package	Exports
@wifo/factory-core	29
@wifo/factory-context	18
@wifo/factory-harness	~16
@wifo/factory-runtime	21
@wifo/factory-spec-review	10
@wifo/factory-twin	~7

Test surface growth

• @wifo/factory-core: 104 → 124 (+scope-project source/init bundling tests + Multi-spec README tests + baseline-reset evidence test + version-pin tests)

Reconciliations worth knowing

• The discoverability gap was the binding constraint, not DoD coverage. Pre-v0.0.7 BASELINE the v0.0.8 plan was DoD-verifier + worktree sandbox + PostToolUse hook + CI publish. The v0.0.7 BASELINE evidence re-ranked entirely: discoverability is the actual ceiling, so DoD-verifier slips to v0.0.9+.
• Slash command is a regular file, not a symlink. Cross-platform reliability: symlinks don't survive npm pack/npm install consistently across macOS/Linux/Windows. The in-repo .claude/commands/scope-project.md IS a symlink for dogfooding (single source of truth), but the scaffolded copy is a plain-file write.
• No retroactive backports. Projects scaffolded before v0.0.8 don't auto-pick-up the new section or slash command — users either re-run factory init or copy the new section by hand. A factory init --upgrade flag is a v0.0.9+ candidate.
• Lockstep bump even for unchanged packages. harness/spec-review/context/twin didn't change in v0.0.8 but bumped to 0.0.8 anyway. Matches the v0.0.5 / v0.0.6 / v0.0.7 publish-coordination pattern; keeps the scaffold's ^0.0.8 deps uniformly resolvable.
• Section content is locked, prose is flexible. The spec locked the section heading (## Multi-spec products) and the structural elements (named commands, code block, auto-install note); the exact wording is implementation-flexible so future doc edits don't trip the structural tests.

---

v0.0.7 — real-product workflow: /scope-project + depends-on + run-sequence

[0.0.7] — 2026-05-02

Theme: real-product workflow. Three deliverables that together collapse the multi-spec-product friction quantified in the v0.0.6 BASELINE run (32 manual interventions per 4-spec product → ~8): a /scope-project slash command, a depends-on frontmatter field, and a factory-runtime run-sequence CLI subcommand. The maintainer now decomposes one product description with one slash command and ships the resulting spec set with one run-sequence invocation; provenance threads the entire product DAG under one factory-sequence record.

Added

• /scope-project slash command (docs/commands/). New canonical source at docs/commands/scope-project.md (in-repo). Takes a natural-language product description and writes 4-6 LIGHT specs in dependency order under docs/specs/<id>.md. First spec ships status: ready; rest ship status: drafting. Each spec populates the new depends-on: frontmatter field. Worked example: docs/baselines/scope-project-fixtures/url-shortener/. Install via cp docs/commands/scope-project.md ~/.claude/commands/scope-project.md.
• depends-on: [<id>, ...] frontmatter field (factory-core). Optional array on SpecFrontmatter. Defaults to []. factory spec lint validates each entry against the new KEBAB_ID_REGEX = /^[a-z][a-z0-9-]*$/ and (with --cwd <dir>) checks that each declared dep file exists under docs/specs/ or docs/specs/done/. Two new lint codes: spec/invalid-depends-on (error), spec/depends-on-missing (warning).
• lintSpecFile(filePath, opts?) helper export (factory-core). Wraps readFileSync + lintSpec with a cwd defaulted to <file>/../... CLI uses this; programmatic callers can keep using lintSpec directly.
• KEBAB_ID_REGEX constant export (factory-core). Canonical kebab-case spec id pattern.
• cross-doc-consistency judge reads declared deps (factory-spec-review). Judge applies() returns true when hasTechnicalPlan || depsCount > 0. New JudgeApplicabilityCtx.depsCount, JudgePromptCtx.deps?: ReadonlyArray<{id, body}>, RunReviewOptions.deps?. CLI auto-loads each declared dep from docs/specs/<id>.md or docs/specs/done/<id>.md; missing → review/dep-not-found warning.
• factory-runtime run-sequence <dir>/ CLI subcommand (factory-runtime). Walks <dir>/*.md, builds the depends-on DAG, topologically sorts via Kahn's (alphabetic tie-break), runs each spec via existing per-spec run() in order. Exit 0 on full converge, 1 on partial / no-converge, 3 on error / cycle / missing dep / sequence-cost-cap.
• runSequence function + SequenceReport type exports (factory-runtime). Public API surface: 19 → 21 names.
• New context record type factory-sequence (factory-runtime). Persisted at sequence start (root, parents=[]); every per-spec factory-run parents at [factorySequenceId] via the new RunArgs.runParents?: string[] arg. factory-context tree --direction down <factorySequenceId> walks the entire product DAG.
• --max-sequence-tokens <n> CLI flag (factory-runtime). Whole-sequence cap on summed agent tokens. Pre-run check: cumulative + nextSpec.maxTotalTokens > maxSequenceTokens aborts before invoking the next spec. Default unbounded (per-spec cap from v0.0.3 still applies).
• --continue-on-fail CLI flag (factory-runtime). Continue running independent specs after a failure; transitive dependents are marked 'skipped' with blockedBy: <first-failed-id>.
• RunReport.totalTokens?: number (factory-runtime). Field-level addition. Computed in-memory from the run's factory-implement-report.tokens. Used by runSequence to accumulate sequence totals.
• Three new RuntimeErrorCode values (factory-runtime): 'runtime/sequence-cycle', 'runtime/sequence-dep-not-found', 'runtime/sequence-cost-cap-exceeded'. Enum count: 10 → 13.
• Two new factory.config.json keys (factory-runtime): runtime.maxSequenceTokens, runtime.continueOnFail. CLI flag > config > built-in default precedence preserved.
• URL-shortener fixture set (docs/baselines/scope-project-fixtures/). Four hand-authored LIGHT spec files demonstrating the canonical /scope-project output shape against the canonical URL-shortener prompt.

Changed

• All six @wifo/factory-* packages bumped to 0.0.7 in lockstep. init-templates scaffold deps bumped from ^0.0.6 to ^0.0.7.

Public API surface

Package	v0.0.6	v0.0.7	Δ
@wifo/factory-core	27	29	+KEBAB_ID_REGEX, +lintSpecFile
@wifo/factory-runtime	19	21	+runSequence, +SequenceReport
@wifo/factory-spec-review	10	10	(field-level on existing types)
@wifo/factory-context	18	18	unchanged
@wifo/factory-harness	~16	~16	unchanged
@wifo/factory-twin	~7	~7	unchanged

Test surface growth

• @wifo/factory-core: 94 → 104 (+8 depends-on schema/lint scenarios + 2 fixture/source structural tests)
• @wifo/factory-runtime: 138 → 161 (+13 sequence + 3 records + 4 runtime.runParents/totalTokens + 3 cli scenarios)
• @wifo/factory-spec-review: 62 → 74 (+8 cross-doc-consistency deps scenarios + 1 review.deps thread + 2 cli dep-loading + 1 review/dep-not-found)
• Workspace total: 446 → 455

Reconciliations worth knowing

• SpecFrontmatter.id is NOT retroactively tightened to match KEBAB_ID_REGEX. The pattern is documented as canonical and enforced ONLY for depends-on entries. Existing specs (and existing third-party specs) may not match. A future v0.0.8+ may add an id-format lint warning for existing-spec ids.
• run-sequence does NOT recurse into <dir>/done/. Specs in done/ already shipped — they are external constraints, not part of the sequence DAG. The cross-doc-consistency reviewer judge handles cross-done/ consistency at review time.
• Sequence-cost-cap is enforced PRE-RUN. Before invoking run() for each spec in topological order, the runtime compares cumulative + nextSpec.maxTotalTokens against maxSequenceTokens. Stricter than post-hoc but prevents one spec from blowing the entire sequence budget on its own.
• Failure cascade is dep-chain-only. A failed spec poisons its TRANSITIVE dependents (via the depends-on chain). Specs with depends-on=[] always run regardless of other failures. --continue-on-fail flips between "run independent roots" and the default "stop on first failure."
• Cascade-blocked status is 'skipped'. Maintains the CI-tooling convention; blockedBy: string field carries the cause.

---

v0.0.6 — v0.0.5.x cluster (4 quality + audit fixes)

Theme: v0.0.5.x cluster shipped together. Four BACKLOG-tracked follow-ups to v0.0.5, bundled because they're all small fixes / quality-of-life improvements that make every subsequent v0.0.6+ workflow cleaner. Three of the four runs hit the very 600s agent timeout that the fourth fix is making configurable — concrete validation that the friction is real.

Added

• --max-agent-timeout-ms <n> CLI flag + RunOptions.maxAgentTimeoutMs?: number (factory-runtime). Default 600_000 (unchanged). Mirrors the v0.0.3 --max-total-tokens pattern exactly: field-level addition, string-label CLI validation (NOT a new RuntimeErrorCode). Wide-blast-radius specs can raise the ceiling explicitly.
• factory.config.json — optional config file at the consumer project root. Specifies defaults for runtime: { maxIterations, maxTotalTokens, maxPromptTokens, noJudge }. Read by factory-runtime run from cwd. Precedence: CLI flag > config file > built-in default. factory init writes one with documented defaults.
• @wifo/factory-spec-review in scaffolded devDependencies (factory-core). factory init now produces a scaffold where factory spec review works on first invocation (previously the dispatch's findPackageRoot walk failed because the package wasn't installed).
• .factory-spec-review-cache in GITIGNORE_TEMPLATE (factory-core). The reviewer cache no longer shows up in git status after a fresh factory init + factory spec review run.

Fixed

• Harness strips surrounding backticks from test: paths and patterns (factory-harness). Recurring spec-authoring pitfall — parseTestLine was passing the literal token (with backticks) to bun test, which never matched any file. Caught twice now (parse-size v1, factory-runtime-v0-0-5). The SPEC_TEMPLATE backtick-guidance BACKLOG entry is now obsolete; spec authors can write either backtick-quoted or bare paths and the harness handles both.
• factory-implement-report.filesChanged is reliable (factory-runtime). Replaces the simple post-run git diff with a pre/post working-tree snapshot (tracked + untracked + content hash). Two failure modes resolved:
  • False negative on new-file-only runs (plain git diff doesn't report untracked files).
  • False positive on pre-run uncommitted changes (the agent gets attributed for files that were already dirty).
    Pre-dirty paths are filtered out — over-attributing the agent is worse than under-attributing.

Changed

• All six @wifo/factory-* packages bumped to 0.0.6 in lockstep (matches the v0.0.5 publish coordination pattern). init-templates scaffold deps bumped from ^0.0.5 to ^0.0.6.
• ROADMAP shift: v0.0.6 was originally themed /scope-project + real-product workflow. That theme moves to v0.0.7; v0.0.6 contains the v0.0.5.x cluster shipped here.

Public API surface

Unchanged across every package. Strictly equal to v0.0.5's surface (27 / 18 / ~16 / 19 / 10 / ~7). All v0.0.6 changes are field-level on already-exported types or internal-only helpers.

Package	Exports
@wifo/factory-core	27
@wifo/factory-context	18
@wifo/factory-harness	~16
@wifo/factory-runtime	19
@wifo/factory-spec-review	10
@wifo/factory-twin	~7

Test surface growth

• @wifo/factory-core: 74 → 77 (+3 init-ergonomics scenarios)
• @wifo/factory-harness: 56 → 60 (+4 backtick-stripping scenarios)
• @wifo/factory-runtime: 131 → 138 (+4 filesChanged + 3 agent-timeout scenarios)
• Workspace total: 446 → 462

Reconciliations worth knowing

• The cluster shipped as v0.0.6, not v0.0.5.1. Spec ids stay as factory-{harness,core,runtime}-v0-0-5-{1,2} (they refer to the v0.0.5 follow-up cluster). The published npm version is v0.0.6 because 4-segment versions like 0.0.5.1 aren't strict SemVer and would be rejected by the registry.
• Three of four runs hit the 600s agent timeout while implementing the cluster. The fourth fix (--max-agent-timeout-ms) is now in for the next run. Net effect: the v0.0.7 work will not hit this — the configurable knob exists.
• Lockstep bump even for unchanged packages. spec-review/context/twin didn't change in v0.0.6 but bumped to 0.0.6 anyway. Matches the v0.0.5 publish-coordination pattern; keeps the scaffold's ^0.0.6 deps uniformly resolvable.

---

v0.0.5 — npm publish + behavior-prior prefix + doc hygiene

Theme: easier to adopt, smarter to use. Every @wifo/factory-* package is now on the public npm registry. factory init scaffolds work outside this monorepo for the first time. The implementer agent gets a stable behavior-prior prompt prefix so prompt caching hits consistently across iterations.

This release was built by the factory itself — factory-runtime ran against three v0.0.5 specs (factory-runtime-v0-0-5, factory-docs-v0-0-5, factory-publish-v0-0-5) with claude -p doing the implementation. Provenance is on disk under .factory-v0-0-5/ for the specs that completed runs.

Added

• npm publish for every package. All six @wifo/factory-* packages (core, context, harness, runtime, spec-review, twin) now ship to npm under v0.0.5 with publishConfig.access: public, full repository/homepage/bugs/keywords/author/license metadata, and per-package LICENSE files.
• Top-level pnpm release script. Gates on typecheck → test → biome check → build → pnpm publish -r --access public. Manual release-gate (not a CI workflow yet).
• # Implementation guidelines prompt prefix in implementPhase. Stable module-level constant IMPLEMENTATION_GUIDELINES emitted between the opening prose and # Spec on every implement spawn. Four behavior priors: state assumptions, minimum code, surgical changes, verifiable success criteria. ≤ 2 KB / ~500 tokens; byte-stable across iterations so claude -p's ephemeral cache hits the same key every iteration.
• PostToolUse hook recipe in packages/core/README.md — opt-in ~/.claude/settings.json block that runs factory spec lint + factory spec review on every Write/Edit to docs/specs/*.md.
• Cross-package review pointers. packages/harness/README.md and packages/runtime/README.md now reference @wifo/factory-spec-review so the reviewer is discoverable from anywhere in the monorepo.
• packages/core/src/publish-meta.test.ts (5 new tests). Asserts per-package metadata, the init-templates dep version (^0.0.5), the caveat sweep across READMEs, and the release-script gating.

Changed

• docs/SPEC_TEMPLATE.md rewritten to match the v0.0.3+ parallel-tree filename convention (docs/specs/<id>.md ↔ docs/technical-plans/<id>.md). Added a "Validating the spec" section listing both factory spec lint and factory spec review. The pre-v0.0.3 single-tree <id>.technical-plan.md convention is gone.
• packages/core/src/init-templates.ts PACKAGE_JSON_TEMPLATE.dependencies bumped ^0.0.4 → ^0.0.5. Scaffold README_TEMPLATE no longer carries the v0.0.4 "monorepo-only" caveat.

Removed

• The v0.0.4 "monorepo-only" caveat is swept from every README + scaffold template (top-level README, packages/core/README.md, packages/spec-review/README.md, examples/{slugify,gh-stars,parse-size}/README.md, init-templates.ts README_TEMPLATE). v0.0.5 is the first release where factory init-generated scaffolds resolve their @wifo/factory-* deps from public npm.
• BACKLOG entry for "implementPhase: behavior-prior prompt prefix" — shipped in this release.

Public API surface

Unchanged across every package. v0.0.5 is metadata + prompt content only — zero new exports, zero rename or removal of existing exports.

Package	Exports
@wifo/factory-core	27
@wifo/factory-context	18
@wifo/factory-harness	~16
@wifo/factory-runtime	19
@wifo/factory-spec-review	10
@wifo/factory-twin	~7

IMPLEMENTATION_GUIDELINES lives in packages/runtime/src/phases/implement.ts but is intentionally NOT re-exported from src/index.ts.

Reconciliations worth knowing

• Coordinated versioning across all six packages. Even packages that didn't change in v0.0.4 (harness was at 0.0.0, twin at 0.0.1) jumped to 0.0.5. Required so the scaffold's ^0.0.5 deps resolve uniformly. Per-package version drift is deferred until there's an operational need.
• pnpm publish -r is manually triggered. v0.0.5 ships the release script and the dry-run-clean tarballs; the actual npm publish is maintainer-driven, not a CI workflow.
• Per-package LICENSE files are copies, not symlinks. Each package gets its own copy so npm pack reliably includes it in every tarball. The root LICENSE is still the source of truth.
• The reviewer's exit-1 condition stays dormant in v0.0.5 — all 5 v0.0.4 judges still default to severity: 'warning'. Promotion is per-judge, post-calibration, in v0.0.5+ point releases.

Notes

This release was the first moneyball run — building v0.0.5 by running v0.0.4's factory-runtime against the three v0.0.5 specs. Three real lessons surfaced and are queued in BACKLOG:

• The harness parses backtick-quoted test paths literally (recurring spec-authoring pitfall).
• SPEC_TEMPLATE.md should warn against the same.
• The 600s per-phase agent timeout is too tight for very-broad specs (the publish spec hit it on iteration 2).

---