Implement task
Build exactly one Task from a Spec folder. The Task file defines the slice,
acceptance criteria, and Verification contract. Standalone execution owns the
full lifecycle; a Roundfix Daemon-assigned turn hands back
implementation-ready work for Daemon Verification and settlement.
Being assigned the Task — by the user, a loop, or a Daemon — is the
authorization to start. Begin immediately: no greeting, no plan-approval
question, no waiting for a "go". Contradictory requirements, or a stale
standalone in_progress status, are the only reasons to pause before
implementing.
Execution modes
Standalone execution and a Roundfix Daemon-assigned turn share the same Task
slice, Result, and evidence requirements, but they have different settlement
owners.
- In standalone execution, the Agent updates Task status, runs every command
in
## Verification, settles the Task from fresh evidence, and commits only
when the user has authorized a commit.
- In a Roundfix Daemon-assigned turn, the Daemon is the sole Task-status
writer. The Agent must not edit status, run the declared
## Verification
commands, claim a terminal verdict, or commit. It may run focused checks,
records implementation and focused-check evidence in ## Result, and hands
back implementation-ready work. The Daemon then runs the complete declared
Verification verbatim before settlement.
- A Daemon Verification command failure releases Verification Capacity before
one Verification Feedback repair turn in the same Agent Session. The Agent
repairs the implementation, updates
## Result, runs focused checks when
useful, and hands back again without running declared Verification or
settling status.
For a Daemon-assigned Task that arrives in_progress, start the Task fresh.
The work-target lock proves that no live Agent owns it. Do not ask how to
resume, and do not normalize or settle its status.
1. Load
Read the assigned task_NN.md, then run the Project Constraint preflight
before changing the Task status to in_progress in standalone execution:
- A completed or archived legacy Spec is exempt from forced constraint
backfill. Leave its artifacts byte-identical and stop rather than
re-executing or rewriting it.
- For an active Spec, require complete
Project Constraints sections in the
PRD and every present TechSpec. Each must account for identifier strategy,
authentication and HTTP, active ADR obligations, and tooling authority as
applicable or not applicable with a reason, with an operative
docs/agents/ source path for every row. Stop without changing Task status
when any row, reason, or source is missing.
- Determine whether the assigned Task creates, edits, renames, moves, or
deletes repository-tooling configuration, scripts, ignore files, plugin
declarations, or version pins. Tooling authorization is not implied by Task
assignment, setup approval, or a generic implementation request.
- Before a tooling Task can start, require express maintainer authorization
and an exact bounded repository-relative file list in the PRD and every
present TechSpec. The mutation allowlist is those exact paths plus the
assigned Task file itself. Missing or conflicting authorization leaves the
Task pending and stops execution.
- Run
git status --short and capture the pre-existing changed paths before
work begins. They remain user-owned and cannot be counted as this Task's
changes.
After the preflight, standalone execution checks status:
pending → set status: in_progress and proceed.
in_progress → a previous session may have died mid-task. Inspect the
worktree and git log for partial work, report what you find, and ask how to
proceed rather than double-building.
completed / failed → stop and report; re-running is an explicit human
decision.
In a Daemon-assigned turn, treat status as Daemon-owned and proceed without
editing it.
- Read the sections of
_prd.md and _techspec.md named in the task's References, plus CONTEXT.md (use its vocabulary in code names and test names) and any referenced ADRs.
- If requirements contradict each other — task vs spec, spec vs ADR — stop and report the conflict. Guessing buries a spec bug inside an implementation.
Skills for the Task
Before editing, read docs/agents/skill-dispatch.md when the repository has
it. Select only the triggers and activation bundles that match the assigned
Task's type, its declared paths and the work it asks for, and load each
selected skill. Never load a skill that matches nothing. Do not load a skill
only a person can start; name it as skipped instead of blocking a Run.
Write the Loaded Skill List in the Task's ## Result on one line:
Skills loaded: <names>, or Skills loaded: none matched. The list is the
Agent's own claim; the Daemon never checks it or uses it to settle the Task.
2. Plan
Build a checklist from Requirements + Acceptance Criteria.
Capture the pre-change signal that proves the task is not yet done — the failing test, the missing behavior, the 404. Without a red starting point you cannot show your change is what turned things green.
Characterization Tasks use the real boundary, within the bounds the Task
declares. When the assigned Task is a characterization Task, exercise the
named external surface and record what the real thing does. Do not substitute
a fake, mock, fixture, or inferred behavior; the record is the boundary
evidence that downstream Tasks may rely on.
Before reaching the boundary, confirm the Task states which target you are
reaching and that you are reaching that one. Read rather than write, or use
the isolated instance the Task creates. Use only the credentials the Task
scopes to that target, honour cancellation, and remove what you created even
when the Task fails. A write against the real boundary happens only where the
Task authorizes that write specifically — recording what a boundary does is
not authority to change it, and an unauthorized write is out of scope
whatever it would prove. If the Task names no target or no bound, stop and
report that rather than choosing one.
3. Implement
- Stay inside the slice. The PRD's Non-Goals and the task's scope are walls, not suggestions — work that belongs to another task goes in a follow-up note, not in this diff.
- A command a requirement names, such as a generator or the owned-skill record command, is part of the work even when Verification runs the same test; its output is never written by hand.
- Tests first at the seams the TechSpec names (they are pre-agreed; a new seam needs the user's sign-off). Typecheck and run the focused tests frequently; save the full suite for the gate.
- Root cause only — no lint/type suppressions, no swallowed errors, no timing hacks. A workaround closes the task and opens a bug.
- For an authorized tooling Task, compare the target path with the mutation
allowlist before every mutation. Never edit
_tasks.md or any other Task
file. If a required path is absent from the authorization, stop and request
a revised Spec instead of widening the list.
4. Standalone verify — the gate
In standalone execution, evidence comes before terminal status in this order:
- Run every command in the Task's
## Verification section verbatim.
- Run the repository's verify pipeline (
make verify, or the build/lint/typecheck/test equivalents this repo documents) when the current execution mode requires local completion evidence.
- Walk Acceptance Criteria one by one: each needs fresh evidence from this session — a command output, a test name that passes, an observed behavior. A green suite is not evidence for a criterion the suite doesn't cover.
For every authorized tooling Task, run a changed-file postflight after the last
edit. Compare the captured baseline with git status --short and
git diff --name-only, accounting for staged, unstaged, and untracked paths.
Every newly changed path must be either one exact authorized path or the
assigned Task file. If any other path appears, make no further mutation, set
status: failed, record the out-of-scope paths in ## Result, and leave the
worktree intact for recovery.
A narrow verification never supports a broad claim. If any standalone check
fails after honest root-cause attempts: set status: failed, record what was
tried in ## Result, and report — a loud failure the scheduler can retry
beats a quiet "mostly done".
5. Standalone record
Append a ## Result section to the task file: what changed (described by behavior, not file lists), commands run with outcomes, evidence per acceptance criterion, and any follow-ups discovered. Tick the Subtasks and Acceptance Criteria checkboxes that the evidence supports. Set status: completed.
Never touch _tasks.md — it owns graph topology, not progress.
6. Standalone commit
- Commit only when the user has explicitly authorized it.
- Stage only this task's files (
git status --short first; unrelated changes stay out).
- One commit per task, Conventional Commits format, task id in the body for traceability (
spec: <slug> / task_02).
- Never push and never open a PR from inside a task — publishing is a separate, explicit action.
7. Daemon-assigned handoff
- Read the assigned Task, PRD, TechSpec,
CONTEXT.md, active ADRs, and bounded
context paths before editing.
- Implement only the assigned slice. Never edit
_tasks.md, another Task
file, or a path outside an authorized tooling allowlist.
- Run focused implementation checks while working when useful, limited to the
packages the change touches. Do not run any command from the Task's
## Verification section, the repository's selected Verification, its
incremental Verification, or the full test suite: the Daemon runs the
declared Verification and the repository Verification at settlement.
- Append or update
## Result with the implementation and focused-check
evidence for every acceptance criterion. Do not use Daemon Verification
evidence that has not run yet.
- Hand back implementation-ready work without editing Task status or claiming
completed, failed, passing Verification, commit readiness, or delivery.
- Never commit, push, or open a pull request. The Daemon runs declared
Verification, writes terminal Task status, and owns the Task commit.
Anti-patterns
- Marking
completed on "should work" — the status field is read by machines that won't double-check.
- Weakening, skipping, or deleting a failing test to get to green.
- Fixing "one more thing" spotted along the way — that's a follow-up, not scope.
- Re-verifying with stale output from earlier in the session; evidence must postdate the last edit.
- Editing other tasks' files or statuses.
- In a Daemon-assigned turn, running declared Verification, editing Task
status, claiming a terminal verdict, or committing.
Fixture values
The credential-assignment Secret Pattern matches an uppercase name ending in
API_KEY, TOKEN, SECRET or PASSWORD, an assignment, and a value of 20 or
more characters containing both a letter and a digit. A fixture value of
letters only, or of fewer than 20 characters, does not match this pattern.
Other Secret Patterns still apply; test files have no exemption.
1---2name: implement-task3description: Execute one task file from docs/specs/<slug>/ end-to-end — ground in the PRD/TechSpec, implement only the slice, and hand back evidence under the execution mode's settlement contract. Starts immediately when assigned, with no confirmation prompt. Use when the user says "run task_03", "execute the next task", "pick up a task", or when an implementation loop or daemon assigns a task file from a spec folder.4metadata:5 category: implementation6 tags: [workflow, coding, agents, testing]7 version: 0.0.68 author: Marcio Altoé9 source: https://skillmd.com/skills/marcioaltoe/implement-task10 upstream: https://github.com/marcioaltoe/skills11version: 0.0.612---1314# Implement task1516Build exactly one Task from a Spec folder. The Task file defines the slice,17acceptance criteria, and Verification contract. Standalone execution owns the18full lifecycle; a Roundfix Daemon-assigned turn hands back19implementation-ready work for Daemon Verification and settlement.2021**Being assigned the Task — by the user, a loop, or a Daemon — is the22authorization to start.** Begin immediately: no greeting, no plan-approval23question, no waiting for a "go". Contradictory requirements, or a stale24standalone `in_progress` status, are the only reasons to pause before25implementing.2627## Execution modes2829Standalone execution and a Roundfix Daemon-assigned turn share the same Task30slice, Result, and evidence requirements, but they have different settlement31owners.3233- In standalone execution, the Agent updates Task status, runs every command34 in `## Verification`, settles the Task from fresh evidence, and commits only35 when the user has authorized a commit.36- In a Roundfix Daemon-assigned turn, the Daemon is the sole Task-status37 writer. The Agent must not edit status, run the declared `## Verification`38 commands, claim a terminal verdict, or commit. It may run focused checks,39 records implementation and focused-check evidence in `## Result`, and hands40 back implementation-ready work. The Daemon then runs the complete declared41 Verification verbatim before settlement.42- A Daemon Verification command failure releases Verification Capacity before43 one Verification Feedback repair turn in the same Agent Session. The Agent44 repairs the implementation, updates `## Result`, runs focused checks when45 useful, and hands back again without running declared Verification or46 settling status.4748For a Daemon-assigned Task that arrives `in_progress`, start the Task fresh.49The work-target lock proves that no live Agent owns it. Do not ask how to50resume, and do not normalize or settle its status.5152## 1. Load5354Read the assigned `task_NN.md`, then run the Project Constraint preflight55before changing the Task status to `in_progress` in standalone execution:56571. A completed or archived legacy Spec is exempt from forced constraint58 backfill. Leave its artifacts byte-identical and stop rather than59 re-executing or rewriting it.602. For an active Spec, require complete `Project Constraints` sections in the61 PRD and every present TechSpec. Each must account for identifier strategy,62 authentication and HTTP, active ADR obligations, and tooling authority as63 applicable or not applicable with a reason, with an operative64 `docs/agents/` source path for every row. Stop without changing Task status65 when any row, reason, or source is missing.663. Determine whether the assigned Task creates, edits, renames, moves, or67 deletes repository-tooling configuration, scripts, ignore files, plugin68 declarations, or version pins. Tooling authorization is not implied by Task69 assignment, setup approval, or a generic implementation request.704. Before a tooling Task can start, require express maintainer authorization71 and an exact bounded repository-relative file list in the PRD and every72 present TechSpec. The mutation allowlist is those exact paths plus the73 assigned Task file itself. Missing or conflicting authorization leaves the74 Task pending and stops execution.755. Run `git status --short` and capture the pre-existing changed paths before76 work begins. They remain user-owned and cannot be counted as this Task's77 changes.7879After the preflight, standalone execution checks `status`:8081- `pending` → set `status: in_progress` and proceed.82- `in_progress` → a previous session may have died mid-task. Inspect the83 worktree and git log for partial work, report what you find, and ask how to84 proceed rather than double-building.85- `completed` / `failed` → stop and report; re-running is an explicit human86 decision.8788In a Daemon-assigned turn, treat status as Daemon-owned and proceed without89editing it.9091- Read the sections of `_prd.md` and `_techspec.md` named in the task's References, plus `CONTEXT.md` (use its vocabulary in code names and test names) and any referenced ADRs.92- **If requirements contradict each other** — task vs spec, spec vs ADR — stop and report the conflict. Guessing buries a spec bug inside an implementation.9394## Skills for the Task9596Before editing, read `docs/agents/skill-dispatch.md` when the repository has97it. Select only the triggers and activation bundles that match the assigned98Task's type, its declared paths and the work it asks for, and load each99selected skill. Never load a skill that matches nothing. Do not load a skill100only a person can start; name it as skipped instead of blocking a Run.101102Write the Loaded Skill List in the Task's `## Result` on one line:103`Skills loaded: <names>`, or `Skills loaded: none matched`. The list is the104Agent's own claim; the Daemon never checks it or uses it to settle the Task.105106## 2. Plan107108- Build a checklist from Requirements + Acceptance Criteria.109- Capture the pre-change signal that proves the task is _not yet_ done — the failing test, the missing behavior, the 404. Without a red starting point you cannot show your change is what turned things green.110- **Characterization Tasks use the real boundary, within the bounds the Task111 declares.** When the assigned Task is a characterization Task, exercise the112 named external surface and record what the real thing does. Do not substitute113 a fake, mock, fixture, or inferred behavior; the record is the boundary114 evidence that downstream Tasks may rely on.115116 Before reaching the boundary, confirm the Task states which target you are117 reaching and that you are reaching that one. Read rather than write, or use118 the isolated instance the Task creates. Use only the credentials the Task119 scopes to that target, honour cancellation, and remove what you created even120 when the Task fails. A write against the real boundary happens only where the121 Task authorizes that write specifically — recording what a boundary does is122 not authority to change it, and an unauthorized write is out of scope123 whatever it would prove. If the Task names no target or no bound, stop and124 report that rather than choosing one.125126## 3. Implement127128- Stay inside the slice. The PRD's Non-Goals and the task's scope are walls, not suggestions — work that belongs to another task goes in a follow-up note, not in this diff.129- A command a requirement names, such as a generator or the owned-skill record command, is part of the work even when Verification runs the same test; its output is never written by hand.130- Tests first at the seams the TechSpec names (they are pre-agreed; a new seam needs the user's sign-off). Typecheck and run the focused tests frequently; save the full suite for the gate.131- Root cause only — no lint/type suppressions, no swallowed errors, no timing hacks. A workaround closes the task and opens a bug.132- For an authorized tooling Task, compare the target path with the mutation133 allowlist before every mutation. Never edit `_tasks.md` or any other Task134 file. If a required path is absent from the authorization, stop and request135 a revised Spec instead of widening the list.136137## 4. Standalone verify — the gate138139In standalone execution, evidence comes before terminal status in this order:1401411. Run every command in the Task's `## Verification` section verbatim.1422. Run the repository's verify pipeline (`make verify`, or the build/lint/typecheck/test equivalents this repo documents) when the current execution mode requires local completion evidence.1433. Walk Acceptance Criteria one by one: each needs fresh evidence from this session — a command output, a test name that passes, an observed behavior. A green suite is not evidence for a criterion the suite doesn't cover.144145For every authorized tooling Task, run a changed-file postflight after the last146edit. Compare the captured baseline with `git status --short` and147`git diff --name-only`, accounting for staged, unstaged, and untracked paths.148Every newly changed path must be either one exact authorized path or the149assigned Task file. If any other path appears, make no further mutation, set150`status: failed`, record the out-of-scope paths in `## Result`, and leave the151worktree intact for recovery.152153A narrow verification never supports a broad claim. If any standalone check154fails after honest root-cause attempts: set `status: failed`, record what was155tried in `## Result`, and report — a loud failure the scheduler can retry156beats a quiet "mostly done".157158## 5. Standalone record159160Append a `## Result` section to the task file: what changed (described by behavior, not file lists), commands run with outcomes, evidence per acceptance criterion, and any follow-ups discovered. Tick the Subtasks and Acceptance Criteria checkboxes that the evidence supports. Set `status: completed`.161162Never touch `_tasks.md` — it owns graph topology, not progress.163164## 6. Standalone commit165166- Commit only when the user has explicitly authorized it.167- Stage only this task's files (`git status --short` first; unrelated changes stay out).168- One commit per task, Conventional Commits format, task id in the body for traceability (`spec: <slug> / task_02`).169- Never push and never open a PR from inside a task — publishing is a separate, explicit action.170171## 7. Daemon-assigned handoff1721731. Read the assigned Task, PRD, TechSpec, `CONTEXT.md`, active ADRs, and bounded174 context paths before editing.1752. Implement only the assigned slice. Never edit `_tasks.md`, another Task176 file, or a path outside an authorized tooling allowlist.1773. Run focused implementation checks while working when useful, limited to the178 packages the change touches. Do not run any command from the Task's179 `## Verification` section, the repository's selected Verification, its180 incremental Verification, or the full test suite: the Daemon runs the181 declared Verification and the repository Verification at settlement.1824. Append or update `## Result` with the implementation and focused-check183 evidence for every acceptance criterion. Do not use Daemon Verification184 evidence that has not run yet.1855. Hand back implementation-ready work without editing Task status or claiming186 `completed`, `failed`, passing Verification, commit readiness, or delivery.1876. Never commit, push, or open a pull request. The Daemon runs declared188 Verification, writes terminal Task status, and owns the Task commit.189190## Anti-patterns191192- Marking `completed` on "should work" — the status field is read by machines that won't double-check.193- Weakening, skipping, or deleting a failing test to get to green.194- Fixing "one more thing" spotted along the way — that's a follow-up, not scope.195- Re-verifying with stale output from earlier in the session; evidence must postdate the last edit.196- Editing other tasks' files or statuses.197- In a Daemon-assigned turn, running declared Verification, editing Task198 status, claiming a terminal verdict, or committing.199200## Fixture values201202The `credential-assignment` Secret Pattern matches an uppercase name ending in203`API_KEY`, `TOKEN`, `SECRET` or `PASSWORD`, an assignment, and a value of 20 or204more characters containing both a letter and a digit. A fixture value of205letters only, or of fewer than 20 characters, does not match this pattern.206Other Secret Patterns still apply; test files have no exemption.