Solve Task - the gated single-chat vertical
One task/feature/bug, six steps, and the user holds the gate between every two. The four twin
skills do the work; this skill owns the chain, the stops, the mode choices, and the state that
survives a compaction or a fresh session. It never designs, builds, or reviews anything itself.
The measurements behind these rules live in references/evidence.md - an audit appendix, not a
run-time load.
When to use
- One task in this chat - not the dispatched multi-agent flow (task-solve-cross), not greenfield, not a one-line edit.
State - the plan file is truth and cursor
- The plan file (
<docs-path>/work/plans/<feature>.md) is the durable truth: the tasks, every stamp this cycle adds (Gated block, Approved +
build mode, Conformance verdict or skipped, Completed), per-task status + evidence - and its
Resume: line is the one cursor (current step, chosen modes, next task), rewritten at EVERY stop and
task tick. On any conflict with memory or the chat, the file wins.
- Handoff notes (the handoff bus,
node .claude/hooks/handoff.js) are the seats' hand-offs in agents mode
(<feature>/<contract_version>/<seat>/<task>), never a second cursor.
On invocation, resume before starting: node .claude/hooks/handoff.js left <feature> (the run's seat notes, by name;
list <feature>/<contract_version> prints them), and read the plan file's stamps and Resume: line. A cycle mid-flight resumes at its cursor - never restart a step whose stamp says it
already passed. A resume paste that lands in a chat still holding the earlier cycle's conversation (no
/clear between) is answered with one line - /clear first, then paste it again - and nothing else.
A NEW cycle starting after a finished one in this same session offers the fresh-session hand-off
in its first ask, after the in-session option. A cycle mid-build looks like:
plan <docs-path>/work/plans/csv.md:
Gated: passed | Approved: 2026-07-16 - mode session
Resume: step 4 BUILD - mode session - next task 2
task 1 DONE (dotnet test green - 4 passed) | task 2 IN_PROGRESS
Invoked with no request and no plan to resume, the one reply is an ask for the task - nothing runs before its answer.
A fresh cycle (no plan stamp) with left <feature> above 0 finds an earlier run's notes under the same name: fold what outlives it into ## Decisions, purge, report stale notes purged: <names>.
The stop contract
A stop IS one AskUserQuestion call: report THREE named fields, then put the next move through the
AskUserQuestion tool - EVERY stop, the plain step-done ones included.
Result: <one line> - <the artifact path>
Progress: <N> of <M> steps
Leftovers: <what this run started and did not finish | none>
none is an answer, an omitted line is not; Markdown-bold (**Result:**) counts, a status table in
place of the names does not.
Every ask marks exactly one option (Recommended), listed first - the move this stop's rule
recommends, the reason in its description. An ask with no mark is malformed: rebuild it before
sending. A multi-select question marks one option too - or asks each judgment item as its own yes/no
question; every option marked reads as a blanket approval. The templates below mark the usual pick;
a stop's own rule may move the mark, never add a second. A step-done stop:
<Step> done - <artifact path>. Continue to <next step>: <the one reason it is next>.
- '<Next step> (Recommended)' - <what it does, in one line>
- 'Route back to <step>' - <the gap or finding that sends it back>
- 'Resume in a fresh session' - required past the fresh-session trigger below
A reply that ASKS is no pick: answer it and end the turn - the stop stays open. The options are
concrete; free text comes through the built-in Other. Where the harness has no such tool, list the same options in plain
text and END THE TURN.
The fresh-session resume. A fresh session (/clear) resumes from the plan file alone, with
zero re-work. Once the cycle has crossed the install's
fresh-session trigger (400,000 tokens on a window of 1M or more, 280,000 on one of 300k up to 1M, 180,000 on any other), spans hours, or resumes after an idle gap past the cache, the fresh-session resume IS one of the next ask's
options - every ask until it is taken or the cycle closes. Below all three (a new cycle's first ask and the compaction stop aside), no ask
carries it and no resume block prints unless the user asks; hours and gaps are the hook's clock, never your guess.
The stop-contract hook injects the two measured numbers (this session's carry per message, a fresh one's cold floor): quote THIS stop's, or say they are unmeasured.
Mark the fresh option only when a hook's fresh-session text in this session says a fresh session pays back, and then it goes first, the in-session option after it unmarked; on Claude Code the stop-contract hook drops it where one does not, and when that leaves one option it refuses the ask - then take the option it names.
Check the trigger before each ask, never from memory. And HONOR the answer: when the user picks it, the turn ENDS with this
block, after the answer - no new work in this chat. The fence holds the command alone: the plan's Resume:
line holds the step.
**Resume in a fresh session** - in <the project root>, type `/clear`, then paste and send:
```text
/task-solve <plan file>
```
Next: <its first move> -> <the steps left>.
Autonomy waiver (AUTO). When the user explicitly asks for a no-stops run ('run all
recommended without asking me'), never self-authorize silently - write <project root>/<data root>/state/approvals/APPROVAL,
first line AUTO - "<their words, verbatim>", say in one line that stops are waived, and take each stop's recommended option; the
pre-commit checkpoint and its receipt still apply. A session past the trigger, or one that already closed a cycle, first puts ONE ask - 'Start AUTO here (Recommended)', then the fresh-session resume - and AUTO starts after its answer. The stamp belongs to the session that dispatches - written when its own decision lands, deleted at its own close; an earlier session's leftover stamp is not consent. references/step-mechanics.md holds the rest.
Size first
State the size in ONE line before anything else - Size: <trivial|small|standard|cross> - <the signal that decided it>, never below the size floor: the full-spec check prints (run it first, on
every size) - then run only that row's steps; it rides the first stop's Result:, Spec: beside it.
The size can be raised mid-run, never lowered - re-checked before a trivial or small close
(references/full-spec.md); a small close's Result: carries review: <reviewer> - <its verdict>.
| Size |
Signals |
Steps that run |
| trivial |
one file, no new dependency, no behaviour a test would see (typo, comment, format, rename inside a file) |
edit -> scoped check -> close. No design, no audit, no stop |
| small |
up to 3 files in one domain, no new dependency, no public contract touched |
plan inline -> build -> task-verify-code inline, no ask -> habits-done-gate -> close. One stop, the close |
| standard |
anything else in one domain |
the full gated vertical below |
| cross |
more than one domain |
the cross-domain orchestrator (/task-solve-cross, the user's command) |
Floor: auth, secrets, input parsing, permissions, a public contract or a migration is never below
standard, whatever the file count.
Full spec - design and audit as one step
On standard, the same check decides this step - references/full-spec.md holds what a FULL spec names and how to run scripts/spec-check.js; any doubt keeps every gate below. State its verdict in one line, Spec: <full|not full> - <path> - <its reason>.
Its path: gated is final. Its path: merged you may raise to gated, never lower.
On path: merged, steps 1 and 2 are ONE step, run by references/full-spec.md. Step 3's read and its habits-execution-strategy load come first (mechanics: read in Result:), then the audit's eight-field block printed as stamped into the plan - never the bare verdict - then ONE approval ask in place of step 3's:
Full spec - designed and audited in one step (<the Passes: line>; <the Gated: verdict>). Build it as planned <in this session | through the seats>: <the mode-fit reason>.
- 'Build as planned - in this session (Recommended)' - <why the tasks fit one chat>
- 'Build as planned - dispatch the agent seats' - <independent tasks that build in parallel>
- 'Not yet - changes needed' - edit the plan, or I ask what to change
The stamps are step 2's and step 3's (Gated: passed, Approved: <date> - mode <session|agents>), and the mark moves to the seats option on step 3's mode-fit rule.
The steps
Each step that names a skill INVOKES it via the Skill tool - again on every new cycle in the same
chat, even when an earlier cycle loaded it. One exception: a capture named in a
close-out line (step 6) is a POINTER for the user to type, never a call this run makes.
DESIGN - run task-design, with habits-clarify loaded first. It writes the plan file -
the artifact, not the chat - with its design rules settled IN it, so step 5 reviews against a plan that already decided them. Stop -
none on a full spec, where step 2 follows in the same step (above).
GATE - run task-verify-plan over the plan file in the mode step 1 ran - no mode ask. It
stamps its eight-field block into the plan: Gated: passed; passed - <N> fixes applied for the
non-judgment fixes it landed in this step, each named in the stop's Result: line; or the gaps,
which route back to step 1 on the user's word. A judgment item is never applied in-step - each
gets its own ask:
Audit judgment: <the item>. Keep or apply?
- 'Keep the plan as written (Recommended)' - the plan stands
- 'Apply it to the plan' - <the audit's reason>
A declined audit is stamped Gated: skipped by user - <their words> - never
blank, never a faked pass. Stop - except on a pass with no open item, where step 3's approval
ask is this stop too ('Not yet' is the route-back).
APPROVE - Read references/step-mechanics.md now - this ask's mode-fit rule and the
mechanics of steps 4-6 are its content, not homework; this stop's Result: line carries mechanics: read as the receipt. Then load
habits-execution-strategy (the Skill tool); its Execution: line is the
mode-fit reason. Then present the gated plan and put the gate through ONE question whose
options each NAME the mode:
The plan is gated. Build it <in this session | through the seats>: <the mode-fit reason>.
- 'Approve - build in this session (Recommended)' - <why the tasks fit one chat>
- 'Approve - dispatch the agent seats' - <independent tasks that build in parallel>
- 'Not yet - changes needed' - edit the plan, or I ask what to change
- 'Resume in a fresh session' - required past the fresh-session trigger above
The mark moves to the seats option when the mode-fit rule picks it for THIS plan, and to the fresh option where the resume rule above marks it. Approval and mode arrive as one answer; a typed Other answer that
omits the mode is re-asked, never defaulted (a mode the invocation named leaves only approve /
not-yet - restate it). Stamp Approved: <date> - mode <session|agents> into the plan file, quoting
the selected answer as the approval words - nothing builds without it.
BUILD - per the approved mode:
- session: run
task-implement - it marks, ticks and resumes each task in the plan file.
- agents: fan the task cards out to the matching
<stack>-implementer seats, each dispatched
exactly as the roster spells it (alfred-code:<seat> where the core plugin carries it), by the
step mechanics' dispatch rules - the approval stamp first, the contract version and the card's
docs anchors in every brief, no isolation, no model: override.
Both modes build to the step mechanics' bar, its build-time stop list included. State what blocks
and what is done, then ONE ask - never 'retry the same command' or 'you run it':
<What blocks task N>. <The in-session route> finishes it here - the task cannot close without it.
- 'Re-scope around it - <the in-session route> (Recommended)' - <what the route does, one line>
- 'Drop the blocked part and close on the rest' - <what stays unbuilt>
- 'Wait for the blocker to clear' - the run stops here until you say go
The harness compacting the context mid-build is a stop too (under AUTO, one line, then continue):
Context compacted during task <N>; the plan file holds every tick. Fresh: <the cold floor>, here: <carry>.
- 'Continue the build here (Recommended)' - on the compacted summary
- 'Resume in a fresh session' - the resume block ends this turn; this chat stops here
Stop at the build's end - and this stop chooses the reviewer for step 5. Its first sentence is
the gate's own verdict line, quoted; a red the run root-caused outside the diff is quoted red,
its cause named, never called green:
<The gate's verdict line, quoted>. Review it <inline | through the seat>: <the reviewer-fit reason>.
- 'task-verify-code in-session (Recommended)' - no dispatch, stays in this context
- 'The <stack>-verifier seat' - isolated eyes, frontmatter model unless you name one
- 'Skip - straight to the done-gate' - stamped as skipped, never the recommendation
The mark moves to the seat when the reviewer-fit rule picks it for the assembled diff.
CONFORMANCE (unless skipped - a skip is stamped Conformance: skipped by user) - INVOKE the reviewer chosen at the step-4 stop, and
re-review each punch-list fix, by the step mechanics' conformance rules. Stamp
Conformance: <verdict> - <reviewer>, <the pass that cleared the last fix delta>; the stop's
Result: carries the review's named fields as it emitted them. Stop - an open finding's ask:
<The review's `Findings:` line, quoted>; open in the request's area: <each BLOCKER / MATERIAL | none>. Fix first?
- 'Fix now (Recommended)' - back to step 4, then the same reviewer on the fix delta
- 'Close now - defer the findings' - each one named in Leftovers
CLOSE - the done gate
(load habits-done-gate - each acceptance criterion, and each Verify or Done-when bullet of a
card the ask cited, demonstrated by a run this session, quoted). Stamp Completed: <date> with the per-task evidence table - it
closes this plan file - and name the ## Decisions ledger by its entry count, never re-pasted. Purge the notes - node .claude/hooks/handoff.js purge <feature> - then the purge count (references/step-mechanics.md).
The close stop's Result: line carries the receipts - Completed <date>, notes purged: <names|none> -
plus Deviations: <each Decisions entry that departs from the Asked: words, one apiece | none>,
under AUTO the user's first sight of them.
Stop - this stop is where the close-out decisions live: anything PENDING (an uncommitted diff, an
unpushed commit, a deferred item, a cross-repo follow-up) goes into the ask; only a cycle with nothing
pending ends on the report alone. An uncommitted diff waits for the user's pick (GIT-PREFERENCES.md):
The feature is done and verified; the diff is uncommitted. Commit it on this machine?
- 'Commit now (Recommended)' - runs `habits-commit-checkpoint` in full, formatter included, then commits
- 'Hold - review the diff first' - nothing is committed; the diff stays as it is
A deferred item or follow-up is a SECOND question in the same AskUserQuestion call, never an
option beside the diff's:
Next: <the deferred item or follow-up>. It opens a NEW cycle in a NEW plan file.
- 'Start it here (Recommended)' - a new cycle in this chat, from step 1
- 'Leave it for later' - named in Leftovers
- 'Start it in a fresh session' - the resume block ends this turn; this context stays behind
An unpushed commit is its own question too (several branches: the merge order):
<sha> on <branch> (<N> commits) is not pushed to <remote><; merge <A> before <B> - why>.
- 'Hold - push later (Recommended)' - nothing leaves this machine
- 'Push now' - the checkpoint's publish half takes this answer as its confirmation
New scope arriving in-chat after Completed: is a NEW cycle in a NEW plan file (a stamped file
is a record, never appended to) - re-enter step 1, or state in one ungated: <why> line that the
work runs outside the cycle; never build it on a casual 'yes, add it'. A change a sibling repo must
follow is handed off as a FILE in THIS repo - a task card under <docs-path>/work/cross-project-tasks/
- and printed as a paste-ready block, by the step mechanics' hand-off rules.
Doc-drift awareness - one line at most in the close report, never auto-run, by the surface
list in
references/step-mechanics.md.
Do not
- Never pass a stop without the user's explicit word, and never approve the plan yourself - the
APPROVE stamp records the user's decision, not yours; an answer that names no build mode
approves nothing.
- Never dispatch a seat the user did not choose at a stop - dispatch is explicit-only house-wide.
- Never keep cycle state only in chat or a note: a stamp, tick or cursor not in the plan file does
not exist.
1---2name: task-solve3description: "Use to run a task through the gated single-chat vertical - 'run the task cycle', 'build this with approvals', 'with my sign-off'. Not a one-line edit."4disable-model-invocation: true5argument-hint: "<the task, feature or bug>"6metadata:7 source: https://skillmd.com/skills/envoydev/task-solve8---910# Solve Task - the gated single-chat vertical1112One task/feature/bug, six steps, and the user holds the gate between every two. The four twin13skills do the work; this skill owns the chain, the stops, the mode choices, and the state that14survives a compaction or a fresh session. It never designs, builds, or reviews anything itself.15The measurements behind these rules live in `references/evidence.md` - an audit appendix, not a16run-time load.1718## When to use1920- One task in this chat - not the dispatched multi-agent flow (task-solve-cross), not greenfield, not a one-line edit.2122## State - the plan file is truth and cursor2324- **The plan file** (`<docs-path>/work/plans/<feature>.md`) is the durable truth: the tasks, every stamp this cycle adds (`Gated` block, `Approved` +25 build mode, `Conformance` verdict or `skipped`, `Completed`), per-task status + evidence - and its26 `Resume:` line is the one cursor (current step, chosen modes, next task), rewritten at EVERY stop and27 task tick. On any conflict with memory or the chat, the file wins.28- **Handoff notes** (the handoff bus, `node .claude/hooks/handoff.js`) are the seats' hand-offs in agents mode29 (`<feature>/<contract_version>/<seat>/<task>`), never a second cursor.3031**On invocation, resume before starting:** `node .claude/hooks/handoff.js left <feature>` (the run's seat notes, by name;32`list <feature>/<contract_version>` prints them), and read the plan file's stamps and `Resume:` line. A cycle mid-flight resumes at its cursor - never restart a step whose stamp says it33already passed. A resume paste that lands in a chat still holding the earlier cycle's conversation (no34`/clear` between) is answered with one line - `/clear first, then paste it again` - and nothing else.35A NEW cycle starting after a finished one in this same session offers the fresh-session hand-off36in its first ask, after the in-session option. A cycle mid-build looks like:3738```39plan <docs-path>/work/plans/csv.md:40 Gated: passed | Approved: 2026-07-16 - mode session41 Resume: step 4 BUILD - mode session - next task 242 task 1 DONE (dotnet test green - 4 passed) | task 2 IN_PROGRESS43```4445Invoked with no request and no plan to resume, the one reply is an ask for the task - nothing runs before its answer.46A fresh cycle (no plan stamp) with `left <feature>` above 0 finds an earlier run's notes under the same name: fold what outlives it into `## Decisions`, purge, report `stale notes purged: <names>`.4748## The stop contract4950A stop IS one AskUserQuestion call: report THREE named fields, then put the next move through the51AskUserQuestion tool - EVERY stop, the plain step-done ones included.5253```54Result: <one line> - <the artifact path>55Progress: <N> of <M> steps56Leftovers: <what this run started and did not finish | none>57```5859`none` is an answer, an omitted line is not; Markdown-bold (`**Result:**`) counts, a status table in60place of the names does not.6162**Every ask marks exactly one option `(Recommended)`, listed first** - the move this stop's rule63recommends, the reason in its description. An ask with no mark is malformed: rebuild it before64sending. A multi-select question marks one option too - or asks each judgment item as its own yes/no65question; every option marked reads as a blanket approval. The templates below mark the usual pick;66a stop's own rule may move the mark, never add a second. A step-done stop:6768```ask69<Step> done - <artifact path>. Continue to <next step>: <the one reason it is next>.70- '<Next step> (Recommended)' - <what it does, in one line>71- 'Route back to <step>' - <the gap or finding that sends it back>72- 'Resume in a fresh session' - required past the fresh-session trigger below73```74A reply that ASKS is no pick: answer it and end the turn - the stop stays open. The options are75concrete; free text comes through the built-in Other. Where the harness has no such tool, list the same options in plain76text and END THE TURN.7778**The fresh-session resume.** A fresh session (`/clear`) resumes from the plan file alone, with79zero re-work. Once the cycle has crossed the install's80fresh-session trigger (400,000 tokens on a window of 1M or more, 280,000 on one of 300k up to 1M, 180,000 on any other), spans hours, or resumes after an idle gap past the cache, the fresh-session resume IS one of the next ask's81options - every ask until it is taken or the cycle closes. Below all three (a new cycle's first ask and the compaction stop aside), no ask82carries it and no resume block prints unless the user asks; hours and gaps are the hook's clock, never your guess.83The stop-contract hook injects the two measured numbers (this session's carry per message, a fresh one's cold floor): quote THIS stop's, or say they are unmeasured.84Mark the fresh option only when a hook's fresh-session text in this session says a fresh session pays back, and then it goes first, the in-session option after it unmarked; on Claude Code the stop-contract hook drops it where one does not, and when that leaves one option it refuses the ask - then take the option it names.85Check the trigger before each ask, never from memory. And HONOR the answer: when the user picks it, the turn ENDS with this86block, after the answer - no new work in this chat. The fence holds the command alone: the plan's `Resume:`87line holds the step.8889````90**Resume in a fresh session** - in <the project root>, type `/clear`, then paste and send:91```text92/task-solve <plan file>93```94Next: <its first move> -> <the steps left>.95````9697**Autonomy waiver (AUTO).** When the user explicitly asks for a no-stops run ('run all98recommended without asking me'), never self-authorize silently - write `<project root>/<data root>/state/approvals/APPROVAL`,99first line `AUTO - "<their words, verbatim>"`, say in one line that stops are waived, and take each stop's recommended option; the100pre-commit checkpoint and its receipt still apply. A session past the trigger, or one that already closed a cycle, first puts ONE ask - 'Start AUTO here (Recommended)', then the fresh-session resume - and AUTO starts after its answer. The stamp belongs to the session that dispatches - written when its own decision lands, deleted at its own close; an earlier session's leftover stamp is not consent. `references/step-mechanics.md` holds the rest.101102## Size first103104State the size in ONE line before anything else - `Size: <trivial|small|standard|cross> - <the105signal that decided it>`, never below the `size floor:` the full-spec check prints (run it first, on106every size) - then run only that row's steps; it rides the first stop's `Result:`, `Spec:` beside it.107The size can be raised mid-run, never lowered - re-checked before a trivial or small close108(`references/full-spec.md`); a small close's `Result:` carries `review: <reviewer> - <its verdict>`.109110| Size | Signals | Steps that run |111|---|---|---|112| trivial | one file, no new dependency, no behaviour a test would see (typo, comment, format, rename inside a file) | edit -> scoped check -> close. No design, no audit, no stop |113| small | up to 3 files in one domain, no new dependency, no public contract touched | plan inline -> build -> `task-verify-code` inline, no ask -> `habits-done-gate` -> close. One stop, the close |114| standard | anything else in one domain | the full gated vertical below |115| cross | more than one domain | the cross-domain orchestrator (`/task-solve-cross`, the user's command) |116117Floor: auth, secrets, input parsing, permissions, a public contract or a migration is never below118standard, whatever the file count.119120## Full spec - design and audit as one step121122On `standard`, the same check decides this step - `references/full-spec.md` holds what a FULL spec names and how to run `scripts/spec-check.js`; any doubt keeps every gate below. State its verdict in one line, `Spec: <full|not full> - <path> - <its reason>`.123124Its `path: gated` is final. Its `path: merged` you may raise to gated, never lower.125126On `path: merged`, steps 1 and 2 are ONE step, run by `references/full-spec.md`. Step 3's read and its `habits-execution-strategy` load come first (`mechanics: read` in `Result:`), then the audit's eight-field block printed as stamped into the plan - never the bare verdict - then ONE approval ask in place of step 3's:127128```ask129Full spec - designed and audited in one step (<the Passes: line>; <the Gated: verdict>). Build it as planned <in this session | through the seats>: <the mode-fit reason>.130- 'Build as planned - in this session (Recommended)' - <why the tasks fit one chat>131- 'Build as planned - dispatch the agent seats' - <independent tasks that build in parallel>132- 'Not yet - changes needed' - edit the plan, or I ask what to change133```134135The stamps are step 2's and step 3's (`Gated: passed`, `Approved: <date> - mode <session|agents>`), and the mark moves to the seats option on step 3's mode-fit rule.136137## The steps138139Each step that names a skill INVOKES it via the Skill tool - again on every new cycle in the same140chat, even when an earlier cycle loaded it. One exception: a capture named in a141close-out line (step 6) is a POINTER for the user to type, never a call this run makes.1421431. **DESIGN** - run `task-design`, with `habits-clarify` loaded first. It writes the plan file -144 the artifact, not the chat - with its design rules settled IN it, so step 5 reviews against a plan that already decided them. *Stop* -145 none on a full spec, where step 2 follows in the same step (above).1462. **GATE** - run `task-verify-plan` over the plan file in the mode step 1 ran - no mode ask. It147 stamps its eight-field block into the plan: `Gated: passed`; `passed - <N> fixes applied` for the148 non-judgment fixes it landed in this step, each named in the stop's `Result:` line; or the gaps,149 which route back to step 1 on the user's word. A `judgment` item is never applied in-step - each150 gets its own ask:151152 ```ask153 Audit judgment: <the item>. Keep or apply?154 - 'Keep the plan as written (Recommended)' - the plan stands155 - 'Apply it to the plan' - <the audit's reason>156 ```157158 A declined audit is stamped `Gated: skipped by user - <their words>` - never159 blank, never a faked pass. *Stop* - except on a pass with no open item, where step 3's approval160 ask is this stop too ('Not yet' is the route-back).1613. **APPROVE** - Read `references/step-mechanics.md` now - this ask's mode-fit rule and the162 mechanics of steps 4-6 are its content, not homework; this stop's `Result:` line carries `mechanics: read` as the receipt. Then load163 `habits-execution-strategy` (the Skill tool); its `Execution:` line is the164 mode-fit reason. Then present the gated plan and put the gate through ONE question whose165 options each NAME the mode:166167 ```ask168 The plan is gated. Build it <in this session | through the seats>: <the mode-fit reason>.169 - 'Approve - build in this session (Recommended)' - <why the tasks fit one chat>170 - 'Approve - dispatch the agent seats' - <independent tasks that build in parallel>171 - 'Not yet - changes needed' - edit the plan, or I ask what to change172 - 'Resume in a fresh session' - required past the fresh-session trigger above173 ```174175 The mark moves to the seats option when the mode-fit rule picks it for THIS plan, and to the fresh option where the resume rule above marks it. Approval and mode arrive as one answer; a typed Other answer that176 omits the mode is re-asked, never defaulted (a mode the invocation named leaves only approve /177 not-yet - restate it). Stamp `Approved: <date> - mode <session|agents>` into the plan file, quoting178 the selected answer as the approval words - nothing builds without it.1794. **BUILD** - per the approved mode:180 - *session*: run `task-implement` - it marks, ticks and resumes each task in the plan file.181 - *agents*: fan the task cards out to the matching `<stack>-implementer` seats, each dispatched182 exactly as the roster spells it (`alfred-code:<seat>` where the core plugin carries it), by the183 step mechanics' dispatch rules - the approval stamp first, the contract version and the card's184 docs anchors in every brief, no isolation, no `model:` override.185 Both modes build to the step mechanics' bar, its build-time stop list included. State what blocks186 and what is done, then ONE ask - never 'retry the same command' or 'you run it':187188 ```ask189 <What blocks task N>. <The in-session route> finishes it here - the task cannot close without it.190 - 'Re-scope around it - <the in-session route> (Recommended)' - <what the route does, one line>191 - 'Drop the blocked part and close on the rest' - <what stays unbuilt>192 - 'Wait for the blocker to clear' - the run stops here until you say go193 ```194195 The harness compacting the context mid-build is a stop too (under AUTO, one line, then continue):196197 ```ask198 Context compacted during task <N>; the plan file holds every tick. Fresh: <the cold floor>, here: <carry>.199 - 'Continue the build here (Recommended)' - on the compacted summary200 - 'Resume in a fresh session' - the resume block ends this turn; this chat stops here201 ```202203 *Stop* at the build's end - and this stop chooses the reviewer for step 5. Its first sentence is204 the gate's own verdict line, quoted; a red the run root-caused outside the diff is quoted red,205 its cause named, never called green:206207 ```ask208 <The gate's verdict line, quoted>. Review it <inline | through the seat>: <the reviewer-fit reason>.209 - 'task-verify-code in-session (Recommended)' - no dispatch, stays in this context210 - 'The <stack>-verifier seat' - isolated eyes, frontmatter model unless you name one211 - 'Skip - straight to the done-gate' - stamped as skipped, never the recommendation212 ```213214 The mark moves to the seat when the reviewer-fit rule picks it for the assembled diff.2155. **CONFORMANCE** (unless skipped - a skip is stamped `Conformance: skipped by user`) - INVOKE the reviewer chosen at the step-4 stop, and216 re-review each punch-list fix, by the step mechanics' conformance rules. Stamp217 `Conformance: <verdict> - <reviewer>, <the pass that cleared the last fix delta>`; the stop's218 `Result:` carries the review's named fields as it emitted them. *Stop* - an open finding's ask:219220 ```ask221 <The review's `Findings:` line, quoted>; open in the request's area: <each BLOCKER / MATERIAL | none>. Fix first?222 - 'Fix now (Recommended)' - back to step 4, then the same reviewer on the fix delta223 - 'Close now - defer the findings' - each one named in Leftovers224 ```2252266. **CLOSE** - the done gate227 (load `habits-done-gate` - each acceptance criterion, and each Verify or Done-when bullet of a228 card the ask cited, demonstrated by a run this session, quoted). Stamp `Completed: <date>` with the per-task evidence table - it229 closes this plan file - and name the `## Decisions` ledger by its entry count, never re-pasted. Purge the notes - `node .claude/hooks/handoff.js purge <feature>` - then the purge count (`references/step-mechanics.md`).230 The close stop's `Result:` line carries the receipts - `Completed <date>, notes purged: <names|none>` -231 plus `Deviations: <each Decisions entry that departs from the Asked: words, one apiece | none>`,232 under AUTO the user's first sight of them.233 *Stop* - this stop is where the close-out decisions live: anything PENDING (an uncommitted diff, an234 unpushed commit, a deferred item, a cross-repo follow-up) goes into the ask; only a cycle with nothing235 pending ends on the report alone. An uncommitted diff waits for the user's pick (`GIT-PREFERENCES.md`):236237 ```ask238 The feature is done and verified; the diff is uncommitted. Commit it on this machine?239 - 'Commit now (Recommended)' - runs `habits-commit-checkpoint` in full, formatter included, then commits240 - 'Hold - review the diff first' - nothing is committed; the diff stays as it is241 ```242243 A deferred item or follow-up is a SECOND question in the same AskUserQuestion call, never an244 option beside the diff's:245246 ```ask247 Next: <the deferred item or follow-up>. It opens a NEW cycle in a NEW plan file.248 - 'Start it here (Recommended)' - a new cycle in this chat, from step 1249 - 'Leave it for later' - named in Leftovers250 - 'Start it in a fresh session' - the resume block ends this turn; this context stays behind251 ```252253 An unpushed commit is its own question too (several branches: the merge order):254255 ```ask256 <sha> on <branch> (<N> commits) is not pushed to <remote><; merge <A> before <B> - why>.257 - 'Hold - push later (Recommended)' - nothing leaves this machine258 - 'Push now' - the checkpoint's publish half takes this answer as its confirmation259 ```260261 New scope arriving in-chat after `Completed:` is a NEW cycle in a NEW plan file (a stamped file262 is a record, never appended to) - re-enter step 1, or state in one `ungated: <why>` line that the263 work runs outside the cycle; never build it on a casual 'yes, add it'. A change a sibling repo must264 follow is handed off as a FILE in THIS repo - a task card under `<docs-path>/work/cross-project-tasks/`265 - and printed as a paste-ready block, by the step mechanics' hand-off rules.266 **Doc-drift awareness** - one line at most in the close report, never auto-run, by the surface267 list in `references/step-mechanics.md`.268269## Do not270271- Never pass a stop without the user's explicit word, and never approve the plan yourself - the272 APPROVE stamp records the user's decision, not yours; an answer that names no build mode273 approves nothing.274- Never dispatch a seat the user did not choose at a stop - dispatch is explicit-only house-wide.275- Never keep cycle state only in chat or a note: a stamp, tick or cursor not in the plan file does276 not exist.