AGENT SKILL

retro

Run a retrospective on a named coding session, on manual invocation only, and file what it finds as backlog tickets in the cycle lane of the project the pass ran in. Reach for it when the user asks for a retro or a retrospective, or asks what a pass should have done differently.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 5
Forks
1

01 前言 Front Matter

name
retro
description
Run a retrospective on a named coding session, on manual invocation only, and file what it finds as backlog tickets in the cycle lane of the project the pass ran in. Reach for it when the user asks for a retro or a retrospective, or asks what a pass should have done differently.

03 SKILL.md 正文

Derived from the retro skill in mattpocock/skills (MIT, see NOTICE.txt), adapted: the output is tickets on the board, not a report to the user.

You suggest improvements to the coding agent's environment — the files, skills, checks and tools it works with — not to the code it wrote. You produce no code and no design; a ticket is the whole output. A retro that files nothing is a valid result; one that files a ticket nobody can act on is not.

Manual only

Retro is a manual standalone pass on a named coding session transcript: the human invokes it (after a sprint, refine, or build pass alike) and decides each time. Nothing ever auto-dispatches or prompts it: not the batch close, not any other skill, not any pass's end.

Read

  1. Load writing-for-agents for the writing style.
  2. The session — opencode api session.list finds the id; opencode api session.message.list --param sessionID=<id> --param limit=20 --param order=asc returns the transcript, tool calls included, oldest first, and it pages by cursor: re-call with --param cursor=<cursor.next from the last response> until a page comes back empty. The envelope is {"cursor": {...}, "data": [...]}: the page is .data, the next call's cursor is .cursor.next, and the empty-page probe is jq '.data|length', never .messages|length, which reads 0 on a full page. --param order=asc goes on the first call only. No offset parameter exists — cursor paging is the only way past the first page, so a single call returns the newest slice only. Two read rules: do not pass order on continuation pages (order + cursor together returns zero messages, silently); redirect each response to a file and jq the file — the CLI byte-truncates a large body when stdout is a pipe (observed cuts at 262144–393216 on a 516 KB body, mid-string, silently). The transcript is the primary source: a finding without a message in it is a guess. No session named → the most recently updated session that is not this one, and say which you took. A read that comes back cut anyway — jq: parse error: Unfinished string at EOF, or a page that fails to parse — means the view is partial: say so, and re-read the missing pages smaller, rather than presenting what you read as the whole transcript.
  3. The environment: the repo's AGENTS.md, its check command, and the board (ticket_list with filter all). A finding an open ticket already covers is a duplicate, not a finding.

Categories

  • Navigation — hard-to-find files, hidden dependencies, a missing pointer.
  • Automated checks — a mistake a check could have caught. Read the repo's own check command first: a check that exists but sits unwired is the finding, and a repo with no guardrail at all is itself a finding.
  • Coding standards — a rule for the reviewer to enforce; a mechanical violation deserves a deterministic check over a written rule.
  • Tool economy — expensive tool calls that could be streamlined.
  • No-ops — steering instructions that change nothing.
  • Information access — information the agent never had.

The bar

A finding becomes a ticket only with all three: an anchor (a file:line, an exact command, or a message that shows the cost), a named cost, and an owner's next move. Not filed: generic advice, a preference with no cost, a restatement of an existing rule, a duplicate, your own run. More than five tickets usually means the bar was lowered — pick the clearest costs and say what you dropped.

The lane

A finding files into the project's lane, on the sprint skill's one derivation (skills/tandem-sprint/SKILL.md, §2): the basename of the directory the retro'd session ran in, with the -wt-<slug> suffix stripped, matched case-insensitively against lane_list and written back in the lane's stored spelling. A Tandem session, in the checkout or in a worktree beside it, files into Tandem: the worktree slug resolves to the project. A finding about another project's files belongs to that project and files into its lane.

Retro is the lane for the findings that belong to no project: a cost with no repo to fix it, a harness rule no project owns. It is not the default, and a finding about a project's own skills, agents or checks never lands there. The origin is separate from the lane and never moves: origin is retro, so the board still says a retro filed it.

File

Dedupe first with ticket_list, filter all and text set to the title words. Then derive lane (the project's, above; Retro when no project owns the finding), call lane_add with name Retro only when filing into Retro and that lane is missing, and file with ticket_add carrying title ("Retro: <one sentence, specific enough to disagree with>"), status backlog, origin retro, lane, and the body in description.

The origin is retro, never assigned. The body is a filing, not a brief: Filed by retro (<session id>)., then Why (the anchor and the cost), What, Done when (only if the session makes it visible). No options, no recommendation. Never edit the code the session touched: a retro that fixes what it finds cannot report honestly on it.

Report

RETRO: <session id> — <n> filed, <m> skipped — <the finding that matters most>

Then the ticket list, one line each, and each skipped candidate with its reason. No questions to the caller, no offer to fix.

The retro's record is the tickets it filed and its final reply: end the reply with the RETRO: line as the summary. File a ticket for every finding; nothing is kept in a capture layer.

04 同仓库更多 Skills

gianniacquisto
gianniacquisto/tandem

writing-for-agents

Writing documents for agents. Use when creating or editing skills, or modifying AGENTS.md or CLAUDE.md.

★ 5
gianniacquisto
gianniacquisto/tandem

tdd

Test-driven development. Use when the user wants to build features or fix bugs test-first, mentions "red-green-refactor", or wants integration tests.

★ 5
gianniacquisto
gianniacquisto/tandem

tandem-tidy

Board hygiene pass: sweeps the open tickets against work that already landed, completes what is done or moot with the evidence in a comment, reports stale claims, archives the done, and reports the oddities. Reach for it when the human says tidy or clean up the board, asks what on the board is stale, or right after a run of merges.

★ 5
gianniacquisto
gianniacquisto/tandem

tandem-team

The roster, which agent fills which role, what to write in `assignee` and `author`, and the minutes rule for the ticket comment log. Reach it when you need to know who owns a pass, what to write as a ticket's assignee or a comment's author, or who the team is.

★ 5
gianniacquisto
gianniacquisto/tandem

tandem-sprint

Triage, build, review and close a batch of tickets in one invocation: orient on the board, draw ready tickets only, run each through `tandem-build`, and settle the lane at the close. Refine and retro are standalone passes: shaping is `skills/tandem-refine/`, learning is `skills/retro/`. Reach for it when the user says run a sprint, asks what the workflow is, or starts a coding pass on this board.

★ 5
gianniacquisto
gianniacquisto/tandem

tandem-refine

Dispatch the refiner to shape a backlog ticket into a buildable brief, the front door for a refine pass, where the human wants to spend time refining rather than building. Reach for it when the user says to refine, spec, or shape a ticket, or asks what needs refining.

★ 5