Work

work — lead implementer

> Requires: the sibling protocols skill (shared protocol masters) and a valid skills.config.json (create with the config skill). Missing either → halt with that exact pointer; never guess.

Pick up the next task, build the right team, drive it to completion, and get it independently verified. Start at the docs router; project constraints come from docs, not from this file.

Read ../protocols/references/docs-convention.md, ../protocols/references/questions.md, ../protocols/references/cross-model.md, ../protocols/references/adversarial-review.md, ../protocols/references/handoff.md, ../protocols/references/reports.md. Validate config FIRST per the config contract — halt on missing/invalid config, naming the field, point at config. Never mutate git (no commit/merge/rebase/push).

> Cross-tool note (Claude Code ↔ Codex). Written in Claude Code idioms. > Under Codex, translate: $ARGUMENTS = the invocation argument; the Agent > tool with subagent_type = = dispatch the matching agent from the > config agents.dir roster (same roles mirrored per host, e.g. > .codex/agents/* alongside .claude/agents/*); "use AskUserQuestion" = ask > in chat. If a role has no host-native agent file, read that role's > definition and apply its checklist inline.

Step 1 — detect current state

Read the router (docs.entry, default docs/README.md) before following any doc path. Resolve the active program/initiative and the next incomplete task from the live documents; never infer current work from archive/. If the user names a task focus, use it as the override; otherwise take the next unchecked task.

Step 2 — load context chain (in order, skip nothing)

1. The active-work doc the router resolves — current initiative, status,

dependencies.

1. The task triad — spec.md, plan.md, tasks.md; find the next

incomplete task (unchecked - [ ]) by its immutable `` ID, never by position or title.

1. Every linked reference/ contract and decisions/ ADR relevant to the

task.

1. The agent roster from config agents.dir (if present) — read each agent

file's description to know which roles apply to this task's domain.

A plan or spec is never evidence that behavior exists. Do not read archive material unless a live doc names a specific provenance need.

Step 3 — interview (plan phase)

Per ../protocols/references/questions.md: triage first — discoverable facts answered from code/docs (never reach the user); technical tradeoffs consult the adversary model, surfacing only on unresolved conflict; product/scope/ preference ALWAYS reach the user. Transport = chat by default (questions.transport); class-c always goes to the user. Cover approach and tradeoffs, ambiguities, edge cases, scope (narrow/expand), and dependencies on other tasks.

Flow mode exception: when invoked by flow (stage 2), grill already ran the interview at stage 1 — SKIP this step and execute the locked triad.

Step 4 — activate teammates

Discover the roster from config agents.dir; read each agent file's description to pick the roles relevant to this task (implementers, reviewers, test/quality, infra). Propose the team (roles + assigned files) to the user and wait for confirmation. Spawn them as subagents/teammates working in the MAIN tree (never a worktree).

If no roster is configured, subagents are unavailable, or the user declines: work as a single agent, but apply each relevant role's checklist inline before considering the task done.

Step 5 — execute

• Project code constraints come from docs, not hardcoded here: read the

conventions prose the config points at (docs.conventions, e.g. docs/CONVENTIONS.md). Honor it exactly.

• Tiered testing via config commands (argv arrays; first use of each

command in a session requires explicit user approval):

• after every change → commands.test (fast targeted suite);
• at task completion → commands.testUnit + commands.lint;
• integration-sensitive work → commands.testAll.

A command absent from config → that tier is skipped; say so, never invent a command. Zero tolerance for failures at every tier: fix before proceeding, before marking done, before status updates.

• Write tests alongside implementation, not after — prefer the tdd skill

(development mode: red-green-refactor) to drive each behavior.

Step 6 — adversarially verify

When the task is complete, the implementing provider must NOT verify its own work. Per ../protocols/references/cross-model.md + ../protocols/references/adversarial-review.md:

• Launch the OPPOSITE provider as a read-only reviewer (Claude impl →

Codex; Codex impl → Claude). Never substitute a same-provider agent. If the opposite CLI is absent/unauthenticated/can't run the proof, stop and ask the user (fix+retry / continue single-model / abort) — never fake a second opinion.

• Give it the spec, the current diff, the applicable convention constraints,

and the exact tier commands. Permit source reads + test runs; forbid edits, commits, implementation.

• Require a structured APPROVED / REVISE verdict with findings, evidence,

commands run, residual risks. Burden of proof is on the diff; "no material objection found" is a valid, logged outcome.

• On REVISE, fix and resume the SAME reviewer session with the new diff.

Loop until agreement or the five-round cap; when disagreement hinges on product intent, present both positions verbatim and the user decides. Persist each verdict to disk as it forms.

Step 7 — documentation

• Flow mode: do NOT mutate docs (single-owner rule). Emit drift_notes

and decisions in the stage report frontmatter (../protocols/references/reports.md / ../protocols/references/handoff.md); the stage-7 librarian pass persists them. Return a compressed summary (≤ ~10 lines).

• Standalone mode: after verification passes, you MAY invoke librarian

(no args) for full closeout — reconcile code/docs, promote knowledge, archive, regenerate nav. Re-run the matching test tier if librarian changed any behavior-governing artifact.

Hand the clean working tree to the user. Never commit, merge, rebase, or push — the user performs all git operations.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: supermodo
• Source: supermodo/skills
• License: MIT

Install and usage instructions live in the source repository linked above.