[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77HH540ZDPMSVNV258T2MJF":3,"skill-same:01M77HH540ZDPMSVNV258T2MJF":50},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":28,"commitSha":29,"skillDirectory":30,"forksCount":31,"repositoryHtmlUrl":32,"repositoryHomepageUrl":33,"ownerHtmlUrl":34,"ownerBlogUrl":35,"ownerDescription":6,"updatedAt":36,"install":37,"resources":49},"01M77HH540ZDPMSVNV258T2MJF","phased-plan",null,"Ceremony rules for shipping a phase of a phased plan — when to release, when to emit a handoff prompt, how far to go on type-strengthening, when to fold or split phases, and the hero-exit rule that prevents plans from becoming open-ended. Load when shipping a phase of a docs\u002F*-plan.md (or equivalent) plan. Use when this named workflow matches the task.","anthony-chaudhary\u002Ffak","anthony-chaudhary","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F18080164?v=4","01M77GSKCAEKV3JS07J40B94DY","Anthony Chaudhary","01M77GSKD26Z7F4FQQD4VA30VX",48,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fphased-plan\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: phased-plan\ndescription: Ceremony rules for shipping a phase of a phased plan — when to release, when to emit a handoff prompt, how far to go on type-strengthening, when to fold or split phases, and the hero-exit rule that prevents plans from becoming open-ended. Load when shipping a phase of a docs\u002F*-plan.md (or equivalent) plan. Use when this named workflow matches the task.\ndisable-model-invocation: true\nuser-invocable: false\nallowed-tools: Read, Edit, Bash\nmetadata:\n  opencode: agent-permission   # #422: model-invocation\u002Fuser-invocable gate re-expressed in opencode.json as `permission.skill: {\"phased-plan\": \"deny\"}` (opencode auto-loads .claude\u002Fskills\u002F** and can't exclude a skill; permission is the real lever)\n---\n\n# Phased-Plan Ceremony\n\nLoad this skill when you finish (or are about to finish) a phase of a phased plan. It governs four things: handoff prompts, releases, plan-shape rules (hero-exit, phase-split, hygiene-queue), type-strengthening. \"The project's release skill\" referenced below is this repo's own `\u002Frelease`.\n\n**Git authorization.** Loading this skill is the user's explicit authorization to run `fak commit` (or explicit-path `git commit`) for the phase-ship commit and for any Phase-0 baseline snapshot the ceremony spells out. The \"never commit unless asked\" default does NOT apply here — committing the phase artifact IS the skill's job. The skill does NOT authorize `git push` or `git tag` directly — those happen via the project's release skill (which has its own authorization), or require explicit user confirmation if invoked outside it. Force-push, history rewrites, branch deletion, and `git reset --hard` always require explicit user confirmation. Use targeted pathspecs only — never `git add -A`\u002F`-u`\u002F`.`.\n\n## Handoff prompts — session-boundary only\n\nEmit a copy-paste handoff prompt **only at a session boundary** — when the user signals they are wrapping up, clearing context, or otherwise ending the session (\"wrapping up\", \"\u002Fclear next\", \"done for the day\", compaction imminent). Do NOT emit one after every phase completion within a live session; the plan doc + issue tracker already carry continuation context for in-session work. Format when you do emit:\n\n```\n--- NEXT-STEP PROMPT (copy into new session) ---\nContinue \u003Cplan-id> (\u003Cdoc path>). Just shipped: \u003Cone-line summary of completed phase>.\nNext phase: \u003Cnext phase id + title>. Goal: \u003C1-sentence goal>. Key files\u002Frefs: \u003Cpaths>.\nStart by reading \u003Cdoc path> and the listed files, then propose a short implementation plan before editing.\n--- END ---\n```\n\n≤6 lines, self-contained, includes plan doc path + next phase id.\n\n## Releases — only on shippable code change\n\nRun the project's release skill **only when the phase produced a user-visible behavior change or modifies shipped code paths**. The commit still lands either way — the release ceremony (version bump, changelog entry, tag, push, release artifact) is reserved for code that will be visible to the operator running the tool.\n\nDo NOT release for:\n\n- Docs-only edits to in-flight plans (plan flesh-out, phase re-sequencing, prose tightening)\n- Proposal \u002F baseline-snapshot \u002F measurement-checkpoint phases that add no runtime code\n- Observation-only \u002F shadow-soak \u002F instrumentation-only phases\n- Single-checkbox issue ticks with no code diff\n- Memory-file or roadmap-index updates\n\nDo release for: new\u002Fchanged CLI flag or subcommand, new\u002Fchanged handler or API endpoint, behavior change in core pipelines, schema changes, bug fixes that affect runtime, new skill or meaningful skill-logic change.\n\nWhen in doubt, skip the release and batch with the next shippable phase. Version numbers should carry signal — raise the bar, not the cadence.\n\n### Pre-release 3-question gate\n\nBefore invoking the release skill, answer:\n\n1. Does the release window contain at least one runtime code-path edit?\n2. Could an operator trigger something from CLI\u002FUI and see different output?\n3. Is the headline a behavior name (not \"plan-doc refresh\", \"phase X captured\", \"ceremony\", \"rollup\")?\n\nThree \"no\"s → docs-only commit on the working branch with no version bump, no tag, no archive, no release entry.\n\n## Progress state — report movement without weakening claims\n\nUse [`docs\u002Fprogress-state-defaults.md`](..\u002F..\u002F..\u002Fdocs\u002Fprogress-state-defaults.md) as the canonical vocabulary. In every plan's **Current state**, report product, evidence, and queue independently; include the next movement. Do not use one overall `KEEP`, `REJECT`, or `HOLD` label as current state.\n\n- A failed or missing runtime receipt changes the evidence axis, not delivered product scope.\n- Mark a validator or correctness slice as `BOUNDED_ACCEPTANCE` only when that named boundary is fully witnessed; state the broader evidence still missing.\n- Award delivery credit for witnessed shipped scope, prepared packets, removed dependencies, launched runs, and reproduced evidence. Award performance credit only after the full applicable claim gate passes.\n- For fak-native performance, preserve exact artifact and quantization identity, `engine=fak-native`, fallback `0`, strict quality, inclusive request-boundary accounting, and a one-variable matched comparison. llama.cpp and MLX remain explicit reference\u002Fcomparison arms only.\n- Apply the native-tuning default in `AGENTS.md`: record candidate source\u002Fconfiguration\u002Fworkload identity and the full configuration delta in the plan. Refresh incompatible reference measurements when objective or load changes; reassess compatible history for SLO-only changes. Justify inseparable bundles and credit only the bundle. Record independent validity\u002Fbudget review before an expensive run, bound to the exact candidate, objective, load, SLO, and planned cost. Candidate changes reopen only affected review and witnesses; existing user authorization, budget, and security limits remain in force.\n- Keep at most three `ACTIVE_PROBE` items. Prefer `READY_TO_RUN`, `DEPENDENCY_ADVANCING`, `QUEUED_BEHIND \u003Cissue>`, or reversible `PARKED_LOW_VALUE` over indefinite `HOLD`. Reserve `EXTERNAL_BLOCK` for exhausted local, fleet, and dependency-clearing routes.\n- Preserve dated historical campaign decisions in **Execution log** even when current state moves to a more actionable queue label.\n\n## Hero-exit rule — close the plan when the hero metric is met\n\nEvery phased plan has a single 1-hop hero metric in its header. When that metric reaches its target, the **next phase queued is the close-out** — not another sub-phase, not a refactor of an adjacent surface, not \"while we're in here.\"\n\nBefore queuing the next phase of a plan whose hero metric you just moved, re-read the metric:\n\n1. **Hero met?** Next phase = plan close-out. In the **same close-out commit**: tombstone the plan-doc, close the umbrella issue, move pending sub-phases to PARK with a \"remaining tail\" label.\n2. **Hero not met?** Continue per the plan's sequencing.\n\nPlans without hero-exit become open-ended consolidation projects that consume slot budget without delivering closure. If work past the hero looks legitimately load-bearing, that means the plan named the wrong hero — fix the metric first (one PR, header only), then continue. Don't ship past a stated hero without auditing it.\n\n## TOMB-at-impl-plus-monitor — soak gates are a lane, not a phase\n\nA plan is tombstone-eligible the moment **implementation ships and a monitor is attached**. \"Monitor\" = either a registered soak watch or a CI invariant that fires on regression. The 100%-phases criterion conflates engineering work with passive evidence-gathering and produces the \"85-95% stuck for weeks\" portfolio pattern.\n\n**Rules:**\n\n1. **Last phase is never \"soak gate + tomb.\"** Those are separate kinds of work. Final engineering phase ships code; soak goes to the soak registry; tomb is administrative.\n2. **TOMB-with-soak-attached is a recognized terminal state.** Plan exits the active queue, soak continues, memory anchor preserves the decision trail.\n3. **A failing soak keeps the plan active.** If a soak's current metric is below its pass-criterion, the plan is *not* tomb-eligible — the risk is real, not theoretical.\n4. **Soak pass-criterion shape:** prefer event-count gates over calendar gates when the system runs hot (e.g. `≥5 events` beats `7d` when traffic is bursty); use composite gates (`metric ≥ X for N consecutive periods`) for noise-prone metrics.\n\n## Phase-split test — fold phases that don't produce intermediate evidence\n\nBefore splitting work into two phases, ask: **\"does phase N+1 need evidence that phase N produces?\"**\n\n- **No** (N+1 is just code chained off N's API surface) → fold them. One phase, one ship-stamp.\n- **Yes** (N+1 reads telemetry, soak result, or measurement that only exists after N ships) → split is justified.\n\nSplitting on operator-fatigue lines is a major producer of fragmentation. The historical default (\"each phase fits one agent session\") is wrong when phase N and N+1 are reversible code chains the same agent could ship together.\n\n**Anti-pattern:** observability-before-guard splits (\"observe\" then \"act\"). When the change is low-risk, bake telemetry into the implementation phase and let the soak prove it.\n\n## Hygiene-umbrella plans are queues, not roadmaps\n\nHygiene-umbrella plans (cleanup, deprecation, tech-debt collections) accumulate open-ended cleanup. Convention:\n\n- One substrate phase at the top (baseline) — fine.\n- Everything after is a **queue**, not pre-numbered phases. Track `inflight items` + `closed in last 7d` count.\n- Drop `% complete` from the plan-doc header (it forces inventing new phase numbers when sibling cleanups surface mid-session).\n- Sibling items batch into one dispatch when they chain on the same substrate.\n\nDistinguishing test: if you can answer \"does the new sub-task need evidence the prior sub-task produces?\" with \"no, it's just more cleanup of the same shape\" → it's a queue item, not a phase.\n\n## One investigation is one plan — do not split it into a series\n\nA single investigation with several suspected causes is **one plan with phases** (Phase 0 baseline → Phase 1..N each a measured fix), not N sibling plans. Sibling plans are justified only when the causes are already proven independent and each has its own hero metric.\n\nRule: **no new plan doc until a Phase-0 baseline names a metric that no existing plan can host.** Before opening any new plan, check whether an existing active or hygiene plan can host the work. Prefer a phase\u002Fqueue-item on an existing plan; open a new doc only when that genuinely fails.\n\n## Phase 0 baselines — register-or-delete invariant\n\n**Optional for low-blast-radius edits.** A phase may skip the Phase 0 baseline freeze when **all three** hold: change is ≤50 LOC, rollback is a single revert, and failure mode is observable post-hoc from existing telemetry. The plan still names the metric and expected direction; it just doesn't pre-freeze a numeric. The measure-then-change axiom is load-bearing for risky changes — wiring substrate, changing dispatch behaviour, touching a hot path — and remains mandatory for those.\n\nWhen a phase DOES produce a measure-then-change baseline, it **must** land in git the same way every time: same commit, registry row (in whatever baseline-tracking file the project uses), no exceptions. Skipping the registry row is the producer for orphan baselines that survive plan-aborts and pile up.\n\n```bash\n# Atomic — one commit, both assets. Never just the dir, never just the row.\nfak sync check\nfak commit \\\n  --path \u003Cbaseline-dir>\u002F --path \u003Cbaseline-registry-file> \\\n  -m \"\u003Cplan>: \u003Cphase> — baseline snapshot (measure-then-change)\"\n```\n\nThe registry row encodes \"this baseline exists, here's what it measures, here's its rebaseline cadence.\" Without it, future tools can't audit the on-disk dir against ground truth — only the operator's memory ties them.\n\n**If the plan aborts:** keep the baseline dir + registry row in git as historical evidence. Annotate the registry row's notes with the abort reason. Deletion is a separate operator decision after the plan has been tombstoned.\n\nNo release ceremony, no version bump — Phase 0 is observation-only by definition (per \"Releases — only on shippable code change\" above).\n\n## Issue-tracker conventions\n\nIf the project uses GitHub\u002FLinear\u002Fetc. for umbrella issues:\n\n- Each active phased plan has one umbrella issue. Tombstoned\u002Fparked plans do not.\n- Pending phases are rendered as `- [ ]` task-list checkboxes inside the issue body so the issue card shows X-of-N progress automatically.\n- When you ship a phase: tick its checkbox.\n- When a plan completes: close the umbrella issue. If the plan tombstones, link the tombstone in the closing comment.\n- When proposing a new plan: create a new umbrella issue, add a link in the plan-doc header.\n\nThe project's CLAUDE.md \u002F AGENTS.md should name the issue-tracker URL and label conventions.\n\n## Type-strengthening — payload-only scope\n\nType the **new payload\u002Fstruct\u002Fenum\u002Fconfig-key the phase itself adds**. Do not scope-creep into surrounding untyped code — that's a separate cleanup, not part of the phase.\n\n- Python: Pydantic model (or dataclass) for any new structured payload crossing a module boundary; `Literal`\u002F`Enum` for closed string sets the phase introduces; return-type annotations on new functions.\n- Go: named struct for any new payload crossing a package boundary; typed enum (`type Foo string` + consts) for any new closed string set; concrete error type when callers of new code will branch on it.\n- Contract lockstep: if the phase adds a new config key or event-payload field, add\u002Fupdate the matching struct in the same commit.\n\nSkip silently when the new thing is genuinely dynamic (third-party shape) or typing would force a breaking change outside the phase's scope.\n","# Phased-Plan Ceremony\n\nLoad this skill when you finish (or are about to finish) a phase of a phased plan. It governs four things: handoff prompts, releases, plan-shape rules (hero-exit, phase-split, hygiene-queue), type-strengthening. \"The project's release skill\" referenced below is this repo's own `\u002Frelease`.\n\n**Git authorization.** Loading this skill is the user's explicit authorization to run `fak commit` (or explicit-path `git commit`) for the phase-ship commit and for any Phase-0 baseline snapshot the ceremony spells out. The \"never commit unless asked\" default does NOT apply here — committing the phase artifact IS the skill's job. The skill does NOT authorize `git push` or `git tag` directly — those happen via the project's release skill (which has its own authorization), or require explicit user confirmation if invoked outside it. Force-push, history rewrites, branch deletion, and `git reset --hard` always require explicit user confirmation. Use targeted pathspecs only — never `git add -A`\u002F`-u`\u002F`.`.\n\n## Handoff prompts — session-boundary only\n\nEmit a copy-paste handoff prompt **only at a session boundary** — when the user signals they are wrapping up, clearing context, or otherwise ending the session (\"wrapping up\", \"\u002Fclear next\", \"done for the day\", compaction imminent). Do NOT emit one after every phase completion within a live session; the plan doc + issue tracker already carry continuation context for in-session work. Format when you do emit:\n\n```\n--- NEXT-STEP PROMPT (copy into new session) ---\nContinue \u003Cplan-id> (\u003Cdoc path>). Just shipped: \u003Cone-line summary of completed phase>.\nNext phase: \u003Cnext phase id + title>. Goal: \u003C1-sentence goal>. Key files\u002Frefs: \u003Cpaths>.\nStart by reading \u003Cdoc path> and the listed files, then propose a short implementation plan before editing.\n--- END ---\n```\n\n≤6 lines, self-contained, includes plan doc path + next phase id.\n\n## Releases — only on shippable code change\n\nRun the project's release skill **only when the phase produced a user-visible behavior change or modifies shipped code paths**. The commit still lands either way — the release ceremony (version bump, changelog entry, tag, push, release artifact) is reserved for code that will be visible to the operator running the tool.\n\nDo NOT release for:\n\n- Docs-only edits to in-flight plans (plan flesh-out, phase re-sequencing, prose tightening)\n- Proposal \u002F baseline-snapshot \u002F measurement-checkpoint phases that add no runtime code\n- Observation-only \u002F shadow-soak \u002F instrumentation-only phases\n- Single-checkbox issue ticks with no code diff\n- Memory-file or roadmap-index updates\n\nDo release for: new\u002Fchanged CLI flag or subcommand, new\u002Fchanged handler or API endpoint, behavior change in core pipelines, schema changes, bug fixes that affect runtime, new skill or meaningful skill-logic change.\n\nWhen in doubt, skip the release and batch with the next shippable phase. Version numbers should carry signal — raise the bar, not the cadence.\n\n### Pre-release 3-question gate\n\nBefore invoking the release skill, answer:\n\n1. Does the release window contain at least one runtime code-path edit?\n2. Could an operator trigger something from CLI\u002FUI and see different output?\n3. Is the headline a behavior name (not \"plan-doc refresh\", \"phase X captured\", \"ceremony\", \"rollup\")?\n\nThree \"no\"s → docs-only commit on the working branch with no version bump, no tag, no archive, no release entry.\n\n## Progress state — report movement without weakening claims\n\nUse [`docs\u002Fprogress-state-defaults.md`](..\u002F..\u002F..\u002Fdocs\u002Fprogress-state-defaults.md) as the canonical vocabulary. In every plan's **Current state**, report product, evidence, and queue independently; include the next movement. Do not use one overall `KEEP`, `REJECT`, or `HOLD` label as current state.\n\n- A failed or missing runtime receipt changes the evidence axis, not delivered product scope.\n- Mark a validator or correctness slice as `BOUNDED_ACCEPTANCE` only when that named boundary is fully witnessed; state the broader evidence still missing.\n- Award delivery credit for witnessed shipped scope, prepared packets, removed dependencies, launched runs, and reproduced evidence. Award performance credit only after the full applicable claim gate passes.\n- For fak-native performance, preserve exact artifact and quantization identity, `engine=fak-native`, fallback `0`, strict quality, inclusive request-boundary accounting, and a one-variable matched comparison. llama.cpp and MLX remain explicit reference\u002Fcomparison arms only.\n- Apply the native-tuning default in `AGENTS.md`: record candidate source\u002Fconfiguration\u002Fworkload identity and the full configuration delta in the plan. Refresh incompatible reference measurements when objective or load changes; reassess compatible history for SLO-only changes. Justify inseparable bundles and credit only the bundle. Record independent validity\u002Fbudget review before an expensive run, bound to the exact candidate, objective, load, SLO, and planned cost. Candidate changes reopen only affected review and witnesses; existing user authorization, budget, and security limits remain in force.\n- Keep at most three `ACTIVE_PROBE` items. Prefer `READY_TO_RUN`, `DEPENDENCY_ADVANCING`, `QUEUED_BEHIND \u003Cissue>`, or reversible `PARKED_LOW_VALUE` over indefinite `HOLD`. Reserve `EXTERNAL_BLOCK` for exhausted local, fleet, and dependency-clearing routes.\n- Preserve dated historical campaign decisions in **Execution log** even when current state moves to a more actionable queue label.\n\n## Hero-exit rule — close the plan when the hero metric is met\n\nEvery phased plan has a single 1-hop hero metric in its header. When that metric reaches its target, the **next phase queued is the close-out** — not another sub-phase, not a refactor of an adjacent surface, not \"while we're in here.\"\n\nBefore queuing the next phase of a plan whose hero metric you just moved, re-read the metric:\n\n1. **Hero met?** Next phase = plan close-out. In the **same close-out commit**: tombstone the plan-doc, close the umbrella issue, move pending sub-phases to PARK with a \"remaining tail\" label.\n2. **Hero not met?** Continue per the plan's sequencing.\n\nPlans without hero-exit become open-ended consolidation projects that consume slot budget without delivering closure. If work past the hero looks legitimately load-bearing, that means the plan named the wrong hero — fix the metric first (one PR, header only), then continue. Don't ship past a stated hero without auditing it.\n\n## TOMB-at-impl-plus-monitor — soak gates are a lane, not a phase\n\nA plan is tombstone-eligible the moment **implementation ships and a monitor is attached**. \"Monitor\" = either a registered soak watch or a CI invariant that fires on regression. The 100%-phases criterion conflates engineering work with passive evidence-gathering and produces the \"85-95% stuck for weeks\" portfolio pattern.\n\n**Rules:**\n\n1. **Last phase is never \"soak gate + tomb.\"** Those are separate kinds of work. Final engineering phase ships code; soak goes to the soak registry; tomb is administrative.\n2. **TOMB-with-soak-attached is a recognized terminal state.** Plan exits the active queue, soak continues, memory anchor preserves the decision trail.\n3. **A failing soak keeps the plan active.** If a soak's current metric is below its pass-criterion, the plan is *not* tomb-eligible — the risk is real, not theoretical.\n4. **Soak pass-criterion shape:** prefer event-count gates over calendar gates when the system runs hot (e.g. `≥5 events` beats `7d` when traffic is bursty); use composite gates (`metric ≥ X for N consecutive periods`) for noise-prone metrics.\n\n## Phase-split test — fold phases that don't produce intermediate evidence\n\nBefore splitting work into two phases, ask: **\"does phase N+1 need evidence that phase N produces?\"**\n\n- **No** (N+1 is just code chained off N's API surface) → fold them. One phase, one ship-stamp.\n- **Yes** (N+1 reads telemetry, soak result, or measurement that only exists after N ships) → split is justified.\n\nSplitting on operator-fatigue lines is a major producer of fragmentation. The historical default (\"each phase fits one agent session\") is wrong when phase N and N+1 are reversible code chains the same agent could ship together.\n\n**Anti-pattern:** observability-before-guard splits (\"observe\" then \"act\"). When the change is low-risk, bake telemetry into the implementation phase and let the soak prove it.\n\n## Hygiene-umbrella plans are queues, not roadmaps\n\nHygiene-umbrella plans (cleanup, deprecation, tech-debt collections) accumulate open-ended cleanup. Convention:\n\n- One substrate phase at the top (baseline) — fine.\n- Everything after is a **queue**, not pre-numbered phases. Track `inflight items` + `closed in last 7d` count.\n- Drop `% complete` from the plan-doc header (it forces inventing new phase numbers when sibling cleanups surface mid-session).\n- Sibling items batch into one dispatch when they chain on the same substrate.\n\nDistinguishing test: if you can answer \"does the new sub-task need evidence the prior sub-task produces?\" with \"no, it's just more cleanup of the same shape\" → it's a queue item, not a phase.\n\n## One investigation is one plan — do not split it into a series\n\nA single investigation with several suspected causes is **one plan with phases** (Phase 0 baseline → Phase 1..N each a measured fix), not N sibling plans. Sibling plans are justified only when the causes are already proven independent and each has its own hero metric.\n\nRule: **no new plan doc until a Phase-0 baseline names a metric that no existing plan can host.** Before opening any new plan, check whether an existing active or hygiene plan can host the work. Prefer a phase\u002Fqueue-item on an existing plan; open a new doc only when that genuinely fails.\n\n## Phase 0 baselines — register-or-delete invariant\n\n**Optional for low-blast-radius edits.** A phase may skip the Phase 0 baseline freeze when **all three** hold: change is ≤50 LOC, rollback is a single revert, and failure mode is observable post-hoc from existing telemetry. The plan still names the metric and expected direction; it just doesn't pre-freeze a numeric. The measure-then-change axiom is load-bearing for risky changes — wiring substrate, changing dispatch behaviour, touching a hot path — and remains mandatory for those.\n\nWhen a phase DOES produce a measure-then-change baseline, it **must** land in git the same way every time: same commit, registry row (in whatever baseline-tracking file the project uses), no exceptions. Skipping the registry row is the producer for orphan baselines that survive plan-aborts and pile up.\n\n```bash\n# Atomic — one commit, both assets. Never just the dir, never just the row.\nfak sync check\nfak commit \\\n  --path \u003Cbaseline-dir>\u002F --path \u003Cbaseline-registry-file> \\\n  -m \"\u003Cplan>: \u003Cphase> — baseline snapshot (measure-then-change)\"\n```\n\nThe registry row encodes \"this baseline exists, here's what it measures, here's its rebaseline cadence.\" Without it, future tools can't audit the on-disk dir against ground truth — only the operator's memory ties them.\n\n**If the plan aborts:** keep the baseline dir + registry row in git as historical evidence. Annotate the registry row's notes with the abort reason. Deletion is a separate operator decision after the plan has been tombstoned.\n\nNo release ceremony, no version bump — Phase 0 is observation-only by definition (per \"Releases — only on shippable code change\" above).\n\n## Issue-tracker conventions\n\nIf the project uses GitHub\u002FLinear\u002Fetc. for umbrella issues:\n\n- Each active phased plan has one umbrella issue. Tombstoned\u002Fparked plans do not.\n- Pending phases are rendered as `- [ ]` task-list checkboxes inside the issue body so the issue card shows X-of-N progress automatically.\n- When you ship a phase: tick its checkbox.\n- When a plan completes: close the umbrella issue. If the plan tombstones, link the tombstone in the closing comment.\n- When proposing a new plan: create a new umbrella issue, add a link in the plan-doc header.\n\nThe project's CLAUDE.md \u002F AGENTS.md should name the issue-tracker URL and label conventions.\n\n## Type-strengthening — payload-only scope\n\nType the **new payload\u002Fstruct\u002Fenum\u002Fconfig-key the phase itself adds**. Do not scope-creep into surrounding untyped code — that's a separate cleanup, not part of the phase.\n\n- Python: Pydantic model (or dataclass) for any new structured payload crossing a module boundary; `Literal`\u002F`Enum` for closed string sets the phase introduces; return-type annotations on new functions.\n- Go: named struct for any new payload crossing a package boundary; typed enum (`type Foo string` + consts) for any new closed string set; concrete error type when callers of new code will branch on it.\n- Contract lockstep: if the phase adds a new config key or event-payload field, add\u002Fupdate the matching struct in the same commit.\n\nSkip silently when the new thing is genuinely dynamic (third-party shape) or typing would force a breaking change outside the phase's scope.\n",{"name":5,"metadata":24,"description":7,"allowed-tools":26,"user-invocable":20,"disable-model-invocation":27},{"opencode":25},"agent-permission","Read, Edit, Bash",true,"main","f565b7b7db63e79649f3a02c08108132a86c5d05",".claude\u002Fskills\u002Fphased-plan",16,"https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak","https:\u002F\u002Fanthony-chaudhary.github.io\u002Ffak\u002Fshowcase.html","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary","https:\u002F\u002Flinkedin.com\u002Fin\u002Fanthonychaudhary","2026-11-15 03:06:00",{"source":8,"skillName":5,"commands":38,"prompt":48,"zipAvailable":27},[39,42,45],{"tool":40,"command":41},"npx","npx skills add anthony-chaudhary\u002Ffak --skill phased-plan",{"tool":43,"command":44},"bunx","bunx skills add anthony-chaudhary\u002Ffak --skill phased-plan",{"tool":46,"command":47},"pnpm","pnpm dlx skills add anthony-chaudhary\u002Ffak --skill phased-plan","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fphased-plan\u002FSKILL.md",[],[51,58,67,74,81,86],{"skillNo":52,"name":53,"nameZh":6,"description":54,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":55,"categories":56,"keywords":57,"publishedAt":19},"01M78V4059KRP2T88YR6Z05MJQ","serve-receipt","On-demand research protocol + deterministic receipt for serving a Hugging Face model on a concrete GPU target — fan out research lanes (model facts\u002Ffiles, engine support matrix, serving recipes, KV\u002Fstate memory math), compute the tiered feasibility verdict with `fak serve-receipt` (never narrated math), and emit a witnessed launch receipt. Use when asked for an \"initial config\", \"serving receipt\", or \"how do I serve \u003Cmodel> on \u003CGPUs>\".","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fserve-receipt\u002FSKILL.md",[],[],{"skillNo":59,"name":60,"nameZh":6,"description":61,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":62,"licenseFileUrl":63,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M77HRGKQV2CJ2K6ZFN4YQ32K","caveman","Ultra-compressed communication mode. Cuts output tokens 65% (measured) by speaking like caveman while keeping full technical accuracy. Supports intensity levels: lite, full (default), ultra, wenyan-lite, wenyan-full, wenyan-ultra. Use when user says \"caveman mode\", \"talk like caveman\", \"use caveman\", \"less tokens\", \"be brief\", or invokes \u002Fcaveman. Also auto-triggers when token efficiency is requested.","MIT","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fsyspromptmmu\u002Fthird_party\u002Fcaveman\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fsyspromptmmu\u002Fthird_party\u002Fcaveman\u002FSKILL.md",[],[],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":19},"01M77HRAY3Y3BE7APN5PVJF7NV","quality-report","Score the repo quality and emit a report card.","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fctxplans\u002Ftestdata\u002Frepo\u002F.claude\u002Fskills\u002Fquality-report\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M77HR9807KHAX13DG9PW2K2Q","ctx-overlay","Compact the context window and prune the memory store when a session grows too long.","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fctxplans\u002Ftestdata\u002Frepo\u002F.claude\u002Fskills\u002Fctx-overlay\u002FSKILL.md",[],[],{"skillNo":82,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M77HR7XQ9X8F2CY9Z4F2ZEZ3","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fctxplanlint\u002Ftestdata\u002Frepo\u002F.claude\u002Fskills\u002Fquality-report\u002FSKILL.md",[],[],{"skillNo":87,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M77HR5BVKCTR1WNJGRBSDNX7","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fctxplanlint\u002Ftestdata\u002Frepo\u002F.claude\u002Fskills\u002Fctx-overlay\u002FSKILL.md",[],[]]