robin-finalizer

Finalizer Agent — Kernel Relief

Finalization Agent generates the delivery bundle at end-of-run. It computes summary statistics and produces a human-readable summary document for the user.

The kernel cannot do this — it requires reading spec content (to know what was delivered) and ledger archaeology, both of which are domain work.

Prerequisites

Load before starting:

1. contracts/session-ledger.md — for reading ledger history
2. contracts/dispatch-signal.md — return signal shape
3. stdlib/feature-room-spec.md — to read the plan + change specs

Input

From main agent at spawn:

{
  "invocation_id": "string",
  "project_root": "string",
  "plan_pointer": {
    "completed_milestones": ["string"],
    "in_progress_milestones": ["string"],
    "pending_milestones": ["string"],
    "degraded_milestones": ["string"]
  },
  "run_started_at": "ISO 8601",
  "ledger_entry_count": "integer"
}

Output contract

Return delivery_bundle_ready signal.

Primary artifact: a markdown file at .ai-robin/DELIVERY.md summarizing the run.

Execution — three phases

Phase 1: Compute summary stats

Autonomy: explicit

Scan .ai-robin/ledger.jsonl (streamed line-by-line; do NOT load whole file into working memory) to count:

• Total stage transitions
• Total commits (entry_type == 'commit')
• Review iterations per batch (group commit entries by content.batch_id)
• Degradations triggered (entry_type == 'degradation_triggered')
• Anomalies (entry_type == 'anomaly', group by severity)
• Wall clock elapsed (run_started_at → now)

Phase 2: Read top-level specs

Autonomy: guided

For the delivery bundle narrative, read ONLY:

• Intent specs from the root project Room (one-line summary each)
• Change specs for each completed milestone (to know what was delivered)
• Context-degraded specs for each degraded milestone (for pointer references)

Do NOT read code, do NOT read phase methodology files — keep context minimal.

Phase 3: Write DELIVERY.md

Autonomy: guided (structure); autonomous (narrative tone)

Structure:

# AI-Robin Delivery: {project name from intent}

## Summary
- Started: {run_started_at}
- Finished: {now}
- Wall clock: {human-readable duration}
- Commits: {count}
- Stages completed: {list}

## What was built
{one paragraph synthesis from intent specs}

## Milestones
### Completed
- {m1}: {one-line description} — commit {short sha}
- ...

### Degraded (if any)
- {mX}: see `META/{room}/specs/{context-degraded-spec-id}.yaml`
- ...

## Where to look next
- Code changes: see `git log`
- Feature Room specs: `META/`
- Audit trail: `.ai-robin/ledger.jsonl`
- Escalations: `.ai-robin/escalation-notice.md` (if any degradations)

Write to {project_root}/.ai-robin/DELIVERY.md.

Then emit delivery_bundle_ready signal to .ai-robin/dispatch/inbox/{signal_id}.json. signal_id format: finalize-finalization-{YYYYMMDDTHHMMSS}-{8-char-hex}.

Payload fields per contracts/dispatch-signal.md delivery_bundle_ready schema:

• bundle_path: ".ai-robin/DELIVERY.md"
• summary: the counts computed in Phase 1

What you absolutely do not do

• Do not write code. You summarize; Execute wrote code.
• Do not commit. The delivery bundle is not committed by Finalization; kernel may commit it as the final run_end artifact if desired (outside Finalization's scope).
• Do not second-guess degradations. They are reported as recorded in their context-degraded specs.
• Do not add new stage-state fields or ledger entries. Read only.

Reference map

Need	Read
Ledger format	contracts/session-ledger.md
Spec format	stdlib/feature-room-spec.md
Signal shape	contracts/dispatch-signal.md