doordash-oss/agentic-orchestrator/implement

Plan Execution

Implement the approved phase plan. You own code, development-time testing, explicitly agent-owned evidence, and the progress handoff. The verification harness owns final contract execution and verification-report.yaml.

Output Files

Start Here

Never create or edit verification-report.yaml or phase_complete. When a testing contract exists, the harness derives the report after it commits your structured root outcome.

1. Read the full phase plan.
2. If {phase_dir}/progress.md exists, read it and resume from ### Where I stopped. Reviewer feedback and the current plan override stale handoff prose.
3. Read testing-contract.yaml if it exists — at {phase_dir}/../testing-contract.yaml for roadmap phases, or {phase_dir}/testing-contract.yaml for cycle layouts (rebase/review-comments, where {phase_dir} is the cycle root). Note only items with owner: agent; final execution of owner: harness items is not your responsibility. If the file does not exist, this feature has no per-iteration machine verification: run every command under the plan's ### Automated Verification yourself each iteration and record the commands and results in your handoff — SUCCESS asserts they pass.
4. Confirm every Task **Repo:** <name> is mounted. If the plan or repo scope is contradictory, call the formal AskUserQuestion control and continue after the answer.
5. Build a dependency graph from Blocked by, shared files, repo tags, and acceptance criteria.

Implementation Rules

• The approved Tasks, acceptance criteria, and Success Criteria define scope.
• For behavior changes, use a red-green-refactor loop: add a focused failing test, verify the expected failure, implement the smallest fix, run the focused test and relevant surrounding suite, then refactor while green.
• Pin discovered bugs with regression tests before fixing them when practical.
• Manual evidence does not replace an automated test for behavior that can reasonably be automated.
• Use stubs only when the plan explicitly requests them. Mark intentional stubs with // STUB(Phase N): <purpose>.
• Do not invent bookkeeping rows or edit the testing contract. If the contract or scope must change, request user input.
• Add comments only for intent, rationale, invariants, or non-obvious tradeoffs.

Task Fan-Out

Parallelize independent non-trivial Tasks. Give each worker the exact Task and acceptance criteria, its repo/file boundary, dependencies, and the focused development test it should use. Sequentialize shared migrations, API decisions, and overlapping edits. Review returned diffs and integrate them before unlocking dependents.

Workers validate the code they change. They do not author final verification results; when a testing contract exists, the harness runs it after handoff. Otherwise, run the plan's Automated Verification before reporting SUCCESS.

Agent-Owned Evidence

For every testing-contract item with owner: agent, write the file named by expected_evidence.path under {iteration_dir}.

• manual_observation: one concise Markdown record covering the contract's consolidated semantic checklist, the observed result, and enough concrete detail for review.
• visual_artifact: the actual image at the specified screenshots/ path.
• behavioral_artifact: one actual trace, recording, or interaction log covering the contract's consolidated primary-journey checklist at the specified behaviors/ path.

Do not write evidence for owner: harness items. Do not invent statuses, counts, transcripts, or waiver claims. If required evidence needs authorization, hardware, or an environment this session lacks, run "$AGENTICO_BIN" report-blocker --contract "{testing_contract_path}" --dir "{iteration_dir}" --items <id,id,...> --capability <name> --reason <text> with the affected contract item ids, then end the iteration with RETRY; the harness pauses on the user gate, which owns waivers and substitutions. Run it as a single bare command with absolute literal paths — no cd, &&, ;, pipes, redirects, command substitution, or shell variables in arguments — or the permission guard refuses it. A row with allow_substitution: true accepts a faithful, labelled substitute. Never create placeholder evidence.

Handoff

At iteration end:

1. Write all required agent-owned evidence files.
2. Write {phase_dir}/progress.md.
3. Run "$AGENTICO_BIN" validate-artifacts --phase implement --role implementer --dir "{iteration_dir}"; fix every reported handoff defect — including evidence gaps: missing, malformed, mis-sized, or byte-identical duplicate captures. This runs the same file-backed checks the post-handoff integrity gate applies — clearing them here avoids losing a whole iteration to a gap you can fix now.
4. Finish the root response with exactly one structured outcome tag from the system prompt. The harness writes phase_complete after validation.

Use this progress.md shape:

# Iteration Progress
 
## Iteration Handoff
 
### Completed this iteration
- <concrete files, functions, or change units landed>
 
### Remaining from the plan
- <unfinished Tasks or Success Criteria; empty when complete>
 
### Where I stopped
<next concrete action, or "Complete">
 
### Gotchas / blockers / in-flight decisions
- <anything the next iteration cannot infer from the plan and diff>
 
## Deferrals
 

deferrals: [] closed_deferrals: []

 
## Iteration State
 
SUCCESS

The Deferrals block must include both keys. Deferrals are work due in a numbered future roadmap phase; permissions, external systems, and human follow-ups are blockers, not deferrals.

Choose exactly one state:

• SUCCESS: implementation, acceptance criteria, development tests, due deferrals, and all owner: agent evidence are complete. When a testing contract exists, the harness performs final contract verification next; otherwise your reported automated-verification runs are the record.
• RETRY: useful implementation progress landed and a concrete in-scope next action is possible in the current environment.

Never emit a bare RETRY for a blocker you cannot act on in this environment (missing credentials, absent hardware): the harness re-dispatches RETRY iterations unchanged. Record the blocker with report-blocker first, or call the formal AskUserQuestion control for a human decision and resume the same iteration after the answer.