--- name: deliver-retro description: Run a retrospective on one or more delivery runs (`deliver` or `flyingnobita-deliver`) and propose improvements to the delivery skill, its runtime checks and role profiles. disable-model-invocation: true --- # Delivery retrospective The user has asked for a **retrospective** on delivery work. It combines a local environment review adapted from Matt Pocock's `retro` with the layer a multi-agent run adds: **coordination** (how the orchestrator ran its agents) and **role split** (whether each piece of work landed with the right role). You propose the smallest change to the delivery system that prevents each loss of time, quality or authority next run, and track whether past fixes held. The retro is read-only toward deliveries. It mutates no run record, pane, worker, worktree or branch. It writes its own [ledger](references/ledger.md), and, after the user's separate choice in step 9, TODO entries in the project's chosen tracker. It edits skills or code only after the user picks candidates for implementation; choosing to log a proposal is not that approval. ## Steps 1. Read and follow the local [environment review](references/environment.md) for the orchestrator's session and delivered repo. Hold its candidates for step 7 rather than presenting them now. Follow the [upstream comparison](references/upstream.md) to identify maintenance differences when an installed upstream copy is available; it is source material, not an instruction to execute `retro`. Done when every environment category has a candidate or an explicit clean or unavailable result, candidates are held for step 7, and the upstream comparison is recorded with its revision or its absence. 2. Resolve the runs. Default to the run this session orchestrated; otherwise the run ID, repo or branch the user names. "Recent runs" or a trend request means every run in the state root for the named repo, newest first. Read [evidence](references/evidence.md) and collect each run's primary sources, then read the [ledger](references/ledger.md) for past retros and their fixes. Done when each run's record, reports, branch history and orchestrator transcript are located, or each missing one is named with the reason. 3. Build a **timeline** per run: stages and user stops, then every task and attempt with its role, effective profile, start, settlement, report, verify result, follow-ups and recoveries. Done when every attempt in the run record appears on the timeline with an outcome. 4. Mark every **rework**: a plan revision, an attempt beyond a task's first, a review iteration beyond the first, a failed gate, a scope change, a worker report the independent run contradicted, and any correction the orchestrator or user made to a role's output. Trace each to its **origin**: the earliest stage and role that had the information to prevent it. Done when every rework has an origin, or is marked unattributable with the reason. 5. Run every check in [checks](references/checks.md) against the timeline and reworks. These complement the local environment review; a finding both cover is reported once, here. Done when every check ID has a finding or an explicit "clean, with evidence". 6. Turn environment and multi-agent findings into candidates, using the ledger's history: - A finding matching a past candidate (same check ID and owning file) is a **recurrence**. If that candidate was implemented, its fix did not hold: raise its severity one level and step its fix type up the ladder below. - An implemented past candidate whose measure the new runs satisfy has **held**; report it in one line. Choose the strongest **fix type** that removes the finding, in this order: 1. **Runtime check**: a deterministic refusal in `deliver-runtime` or `deliver-orca-preflight` (the deliver skill's `scripts/`), with a test in the delivery repository's `tests/`. Use for a **mechanical** finding (a missing field, an unverified state passing a gate, a duplicate dispatch) that recurs, or whose first occurrence passed a gate. The candidate states the refusal condition, the command that refuses, a test case built from this run's evidence, and any legitimate run it could wrongly refuse. 2. **Skill text**: a sharper step, completion criterion or role brief in the file that owns the behaviour (table below). The default for a first mechanical finding that passed no gate. 3. **Profile**: a role's harness, model, effort or mode. 4. **Project rule**: the delivered repo's `AGENTS.md` `## Delivery` section, `docs/agents/delivery.toml`, or a repo check from step 1. 5. **Dependency**: an upstream `mattpocock-*` skill; record the contract change in [dependencies](../deliver/references/dependencies.md). | Behaviour | Owning file under `skills/deliver/` | |---|---| | Stops, stage order, orchestrator vs role edits, ticket split rule | `SKILL.md` | | Brief, grilling depth, size route | `references/intake.md` | | Role briefs, build and red-check gates, ticket graph merges | `references/pipeline.md` | | Review classification and repair loop | `references/review.md` | | Frontier, questions, human work, partial runs | `references/scheduling.md` | | QA guardrails and method | `references/qa.md` | | Landing and publication | `references/ship.md` | | Concurrency, landing and role settings | `references/configuration.md` | | Native operations, waits, permissions, recovery | `adapters/.md` | | Herdr default profiles | `references/herdr-profiles.toml` | | Herdr project role profiles | the delivered repo's `docs/agents/delivery.toml` `[hosts.herdr.roles.]`, changed through `deliver-setup` (or `flyingnobita-deliver-setup`) | | Paseo shared profiles | Paseo profile configuration (daemon / MCP) | | Orca default profiles | `references/orca-profiles.toml` | | Orca project role profiles | the delivered repo's `docs/agents/delivery.toml` `[hosts.orca.roles.]`, changed through `deliver-setup` (or `flyingnobita-deliver-setup`) | Done when every candidate names one fix type, one owning file and a **measure**: the observable in the next run's evidence that shows it worked. 7. Present the candidates in severity order: 1. **Authority**: a gate passed on unverified evidence, a role or the orchestrator acted beyond its approval, or a stop was skipped. 2. **Correctness**: a defect shipped, or was caught only by QA, CI or the user. 3. **Rework**: loops and revisions, ranked by elapsed time lost. 4. **Waste**: idle time, duplicated legwork, oversized context. Give each candidate as: title; check ID; evidence (run ID, attempt or report path, timestamp or transcript line); origin; recurrence count; fix type and owning file with the proposed wording or check; measure. Merge the held environment candidates into the same order, labelled by their category. Then give the trends from the ledger: recurring findings, fixes that held, and the gate yield and profile fit tables from checks R7 and R9. Display duration columns as hours and minutes (`Hh Mm`), for example `337` minutes as `5h 37m` and `22` minutes as `0h 22m`. Round fractional minutes once to the nearest whole minute before splitting into hours and minutes. Preserve qualifiers such as `≥`; show unknown time as unknown, never zero. Keep numeric minutes in the ledger for calculations. End with the clean checks in one line. 8. Write this retro's ledger entry, with every candidate `proposed`. Done when the entry exists and the user has the candidate list. 9. End the retro with one offer to store its remaining TODOs in the project's chosen tracker. Name the destination and concrete candidate IDs/titles; offer all, selected IDs, or none. Include already accepted but unfinished work, and exclude completed, declined and duplicate items. If no TODOs remain, state that instead of offering empty work. If the destination is unknown, ask for it with the same offer; preserve an explicit `none` choice. The user may answer later. Until then keep the candidates in the ledger without writing external tracker entries or assuming consent from silence. On acceptance, use the shared [outstanding-work procedure](../deliver/references/outstanding.md): search for existing work, create/update the selected entries, read them back and return their links. Record each candidate's `tracker_refs` in the ledger. Logging proposed candidates keeps them `proposed`; implementation approval remains separate. If a write fails, retain the TODOs in the ledger and name the unsaved items. Done when the offer is presented and its pending, declined or verified-saved outcome is explicit. When the user picks candidates, update their ledger status. Implement picked candidates in the delivery repository under `skills/` or `tests/`, or the delivered repo for project rules, with a CHANGELOG entry, commit following repository commit rules (or a configured custom commit skill), and record the commit in the ledger. Publication is a separate request.