senior-execute

Prompt file imported from Nom-nom-hub/senior-agent (.codex/prompts/senior-execute.md). Fill in {{arguments}} before use. Copyright stays with the author.

Arguments

{{arguments}}

You MUST consider the user input before proceeding (if not empty).

What this stage is

Mechanical execution of tasks.md — with the senior's discipline: tests first, small increments, clean state between units, honesty about failures. The judgment is already spent in design; this stage's job is to not waste it.

Execution steps

1. State check: senior state --json. Requires design.md and tasks.md. Read them plus .senior/principles.md (governing) and .senior/work/<slug>/decisions-refs if the design cites decisions. If tasks.md is missing or stale relative to design.md, STOP and re-derive via /senior.design.
2. Pre-flight:
   • Confirm the working tree is clean enough to produce attributable commits (git status short report; warn the user about unrelated dirty files, do not stash without asking).
   • Confirm the branch situation matches scope.md's rollback plan; create the work branch if scope.md specifies one and it does not exist.
   • Record the baseline: run the existing test suite + typecheck + lint BEFORE any change, and save results. If the baseline is red, report it and get user consent before proceeding — you can't detect your own breakage on a red baseline.
3. Execute tasks in order, per discipline:
   • Test-first contract: for each task marked as a contract/failing-test task, write the test, run it, confirm it FAILS for the right reason, then implement until green.
   • Small increments: implement one task, run the nearest verification (affected tests
     • typecheck), then mark [X] in tasks.md with a one-line commit if principles.md's commit policy asks for per-task commits.
   • [P] parallel tasks may batch, but only across disjoint files, and never batch a test task with its implementation newness check.
   • Halt rules: a failing test, a typecheck error, or a lint error introduced by this work halts the pipeline until fixed. Fix forward within the task; if the fix reveals a design flaw, STOP, record the finding, and route back to design — silently hacking around a bad design is the signature non-senior move.
   • Deviation log: any change that leaves design.md's stated paths/behavior gets a row in a "## Deviations" section appended to design.md (<what> <why> <how verified>) AND its rationale checked against decisions.md.
   • Scope guard: work not traceable to a task is not done. Ideas encountered mid-execution go to .senior/backlog.md, never into this diff.
4. Per-increment verification, minimum: nearest tests + typecheck. No "it should pass".
5. Completion pass: run the FULL test suite + typecheck + lint once all tasks are [X]; verify against design.md section by section; ensure no leftover TODOs or dead code point to unexecuted intentions. Report: tasks done, full-suite status, deviations count, and next command /senior.review <slug>.

Rules

• tasks.md is the execution log: mark [X] only when genuinely complete AND verified; never batch-check at the end.
• No skipping the failing-first confirmation on contract tasks (write it down if the environment legitimately can't run tests — that goes in the Deviations section).
• Never weaken an existing test's assertions to make work pass. If a test needs changing, the change is justified in the Deviations log and the assertions preserved or strengthened.

Done When

• All tasks [X], each verified at the moment of completion
• Contract tests were confirmed failing first (or deviation recorded + explained)
• Full test suite + typecheck + lint green (or explicitly reported as blocked, with why)
• Deviations section (if any) enumerates every departure from design with rationale
• Baseline preserved: nothing unrelated broken, tree attributable per commit
• User told: stage complete, next /senior.review <slug>