agentleFS
Yesterday-AIYesterday-AI/ytstack/skills/lane-retro/SKILL.md

lane-retro

Skill2 starsChanged 34 days agoMITView on GitHub

Looks at a sample of finished lanes while coordinated work is still running and proposes checks and pointers for what went wrong in them. Use when several lanes finished, when a lane went wrong, when the same finding came back in a review, or when the user says lane retro, why does this keep happening, what do the lanes keep getting wrong. Not the retrospective of a finished milestone (retro) and not a review of one change (code-review).

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Step 4 -- Read four things per lane: Only these; not the transcript, not the diff, not the rest of the repository.

    line 139
  2. Distinctive:

    Step 8 -- Ask once: No candidates: skip to Step 10 (Upstream).

    line 198

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

lane-retro

A coordinator sees the same mistake in the third lane and is about to write a fourth sentence about it. This skill looks at a few finished lanes instead, in six fixed categories, and ends in a list of candidates: each one a check or a pointer, never a new rule in prose. The human accepts or skips them in one reply. Accepted ones become tasks. Nothing is implemented here.

Anti-Pattern: "I'll put it into the brief so that the next worker knows"

The next worker gets a longer brief and makes the same mistake, because the sentence competes with everything else it has to hold. A mistake a machine can detect gets a check, and the sentence it replaces is deleted. A search the worker had to make gets a link. A rule that was written down and violated anyway is the strongest case for a check, not for bold type.

Checklist

You MUST create a TodoWrite task for each of these items and complete them in order:

  1. Run preamble -- finished lanes, which went wrong, which were sampled before
  2. HARD-GATE -- .ytstack and at least one finished lane
  3. Pick the sample -- about one finished lane in five, the ones that went wrong first
  4. Read four things per lane -- brief, report, review findings, row; nothing else
  5. Examine in six categories -- one line per lane and category, with evidence
  6. Form candidates -- each a check or a pointer that replaces a named line or a search
  7. Write .ytstack/retro/LANES-<date>.md
  8. Ask once -- all candidates numbered, a recommendation each (only if there are candidates)
  9. Turn accepted candidates into tasks -- a task line or a quick task; implement nothing
  10. Upstream -- what of these lanes concerns ytstack itself, handled per UPSTREAM_FEEDBACK, recorded under ## Upstream
  11. Report + return

Process Flow

digraph lane_retro {
    "Run preamble" [shape=box];
    "Finished lane?" [shape=diamond];
    "Sample: four things per lane, six categories" [shape=box];
    "Candidates?" [shape=diamond];
    "Write LANES-<date>.md" [shape=box];
    "Ask once" [shape=box];
    "Accepted -> task or quick task" [shape=box];
    "Upstream per setting" [shape=box];
    "Stop: nothing to look at" [shape=doublecircle];
    "Report + return" [shape=doublecircle];

    "Run preamble" -> "Finished lane?";
    "Finished lane?" -> "Stop: nothing to look at" [label="no"];
    "Finished lane?" -> "Sample: four things per lane, six categories" [label="yes"];
    "Sample: four things per lane, six categories" -> "Write LANES-<date>.md" -> "Candidates?";
    "Candidates?" -> "Upstream per setting" [label="none"];
    "Candidates?" -> "Ask once" [label="yes"];
    "Ask once" -> "Accepted -> task or quick task" -> "Upstream per setting";
    "Upstream per setting" -> "Report + return";
}

Preamble

# ytstack resolver (M016): checkout = where this shell runs, via the shared lib.
_LIB="${CLAUDE_SKILL_DIR}/../../hooks/lib/git-ctx.sh"
if [ -f "$_LIB" ]; then
  . "$_LIB"; yt_git_ctx "$PWD"; yt_resolve_yt_dir >/dev/null
  _PROJECT_DIR="$YT_TOPLEVEL"; _YT_DIR="$YT_DIR"; _YT_SCOPE="$YT_SCOPE"; _IS_GIT="$YT_IS_GIT"; _PROJECT_KEY="$YT_REPO_KEY"
else
  # Host did not substitute CLAUDE_SKILL_DIR (e.g. Codex adapter): project scope only.
  _IS_GIT=$([ "$(git rev-parse --is-inside-work-tree 2>/dev/null)" = true ] && echo yes || echo no)
  _PROJECT_DIR=$(git rev-parse --show-toplevel 2>/dev/null || pwd -P)
  _YT_DIR=""; _YT_SCOPE=none; _PROJECT_KEY=""
  [ -d "$_PROJECT_DIR/.ytstack" ] && { _YT_DIR="$_PROJECT_DIR/.ytstack"; _YT_SCOPE=project; }
fi
cd "$_PROJECT_DIR" 2>/dev/null || true
_PROJECT_SLUG=$(basename "$_PROJECT_DIR")
_HAS_YTSTACK=$([ -n "$_YT_DIR" ] && echo yes || echo no)

_NI=false
if [ -n "${YTSTACK_NON_INTERACTIVE:-}" ] || [ -n "${OPENCLAW_SESSION:-}" ] || [ -n "${CLAUDE_AGENT_TEAM_MEMBER:-}" ]; then _NI=true; fi
_NL='
'
# Lanes an earlier lane retro looked at: the "sampled:" line of its frontmatter.
_SAMPLED=""
for _p in "$_YT_DIR"/retro/LANES-*.md; do
  [ -n "$_YT_DIR" ] && [ -f "$_p" ] || continue
  _SAMPLED="$_SAMPLED$_NL$(sed -n 's/^sampled: *//p' "$_p" | tr ',' '\n' | sed 's/^ *//; s/ *$//')"
done
# Finished lanes: merged ones, and blocked ones (those went wrong).
_FIN=""
[ -f "$_YT_DIR/LANES.md" ] && _FIN=$(awk -F'|' '/^\|/ { n++; if (n <= 2) next
    for (i = 2; i <= 11; i++) gsub(/^[ \t]+|[ \t]+$/, "", $i)
    if ($4 == "merged" || $4 == "blocked") print $2 "\t" $4 "\t" $10 "\t" $8 }' "$_YT_DIR/LANES.md")

echo "HAS_YTSTACK: $_HAS_YTSTACK"
echo "PROJECT_SLUG: $_PROJECT_SLUG"
echo "BRANCH: $(git branch --show-current 2>/dev/null || echo unknown)"
echo "YTSTACK_NON_INTERACTIVE: $_NI"
_NFIN=0; _NNEW=0
echo "FINISHED_BEGIN"
while IFS="$(printf '\t')" read -r _lane _st _round _pr; do
  [ -n "$_lane" ] || continue
  _NFIN=$((_NFIN + 1)); _mark=""
  case "$_round" in ''|*[!0-9]*) _round=0 ;; esac
  if [ "$_st" = blocked ] || [ "$_round" -ge 3 ]; then _mark="went-wrong"; fi
  case "$_NL$_SAMPLED$_NL" in
    *"$_NL$_lane$_NL"*) _mark="${_mark:+$_mark, }sampled" ;;
    *) _NNEW=$((_NNEW + 1)) ;;
  esac
  echo "$_lane | $_st | round $_round | PR $_pr | ${_mark:--}"
done <<EOF
$_FIN
EOF
[ "$_NFIN" -gt 0 ] || echo "no finished lanes"
echo "FINISHED_END"
echo "FINISHED: $_NFIN"
echo "SAMPLE_SIZE: $(( (_NNEW + 4) / 5 ))"
_TODAY=$(date -u +%Y-%m-%d)
echo "RETRO_FILE: $_YT_DIR/retro/LANES-$_TODAY.md"
echo "RETRO_FILE_EXISTS: $([ -f "$_YT_DIR/retro/LANES-$_TODAY.md" ] && echo yes || echo no)"
# Upstream feedback (M033): the effective setting; the rules are docs/ux/upstream-feedback.md.
_UP="${CLAUDE_SKILL_DIR}/../../bin/ytstack-upstream"; [ -f "$_UP" ] || _UP=$(command -v ytstack-upstream 2>/dev/null)
_UPM=""; [ -n "$_UP" ] && _UPM=$(bash "$_UP" -C "$_PROJECT_DIR" mode 2>/dev/null | sed -n 's/^mode: //p')
echo "UPSTREAM_FEEDBACK: ${_UPM:-unavailable}"
echo "UPSTREAM_TOOL: ${_UP:-unavailable}"

Procedure

Step 2 -- HARD-GATE. HAS_YTSTACK: no: stop, "no .ytstack/ here". FINISHED: 0 and the user named no lane: stop, "no finished lane to look at yet".

Step 3 -- Pick the sample. Take SAMPLE_SIZE lanes from the list between FINISHED_BEGIN and FINISHED_END (about one finished lane in five), in this order: lanes marked went-wrong (blocked, or a third review round); lanes whose report said Status: partial or Status: blocked, or whose verification failed; then the most recently merged. Skip lanes marked sampled. A lane the user names is always in the sample.

Step 4 -- Read four things per lane. Only these; not the transcript, not the diff, not the rest of the repository. If one of the four does not exist for a lane, say so in the file: that is itself evidence for "access to information".

  • its brief: "${CLAUDE_SKILL_DIR}/../../bin/ytstack-brief" render --slice <lane> (--task for a task lane) prints what the worker was given;
  • its report: the labelled lines the coordinator kept, or the lane's M###-S##-T##-SUMMARY.md files;
  • its review findings: the ## Review lines in its task plans and the findings of its review (in the pull request, read from the host for this purpose; say when);
  • its row in LANES.md.

Step 5 -- Examine in six categories. The same six for every lane, in this order:

CategoryThe questionWhere the evidence is
navigation pointersDid the worker find what it needed, or did it search?Deviations: and Not checked: of the report; a file the diff needed and the brief did not name
automated checksWas a mistake caught by a machine, by a reviewer, or by nobody?Checks run: against the review findings
review standardsWas the finding a rule that exists, a rule that is missing, or taste?the finding against AGENTS.md, CONTRIBUTING.md, the project's review policy
instruction lines that change nothingWhich rule was written down and violated anyway?the finding and the line of the instruction file it repeats
tool economyWhich calls only polled or read the same thing again?repeated Checks run: on one commit; a status that was fetched in a loop
access to informationWhat could the worker not reach?Blockers: and Not checked:

A category without evidence in a lane gets "nothing found", not a guess.

Step 6 -- Form candidates. A candidate needs evidence: the lane, the file and the line of the report or the review. Its proposal is exactly one of two kinds: a check (a lint rule, a hook, a test, a CI job) that replaces a named instruction line, which is deleted when the check exists; or a pointer (a link in the plan's ## Files or the brief's reading list, or in the area file of the technical spec) that replaces a search. What decides between them:

  • A mechanical mistake never gets new prose as its answer.
  • A rule that exists and was violated is a candidate for a check and for deleting the line.
  • Reviewers enforce standards, because the implementer carries the most context pressure: a standards candidate goes to the review brief (what a review lane is told), not to the implementer's.
  • A finding that is taste, or that no check and no pointer would have prevented, is no candidate. It is listed under "Looked at, no candidate".
  • What a worker could not reach at all is a question for the human (.ytstack/questions/), not a candidate.

Step 7 -- Write the file. Path: RETRO_FILE, candidates in the order of the six categories. RETRO_FILE_EXISTS: yes: append a second # Lane retro block with its time and add the lanes to sampled:; never rewrite the earlier one.

---
kind: lane-retro
written: {UTC}
sampled: {lane}, {lane}
of_finished: {FINISHED}
---

# Lane retro {UTC}

## Candidates

### C1 -- {category}: {the mistake in one sentence}

- Evidence: {lane}, `{file}:{line}` -- {what the report or the review says there}
- Proposal: {check | pointer} -- {what exactly, and where it lives: CI, a hook, the test suite, the review brief, the plan's Files, spec/{area}.md}
- Replaces: `{instruction file}:{line}` | the search for {what}
- Recommendation: {accept | skip} -- {one reason}
- Decision: open

## Looked at, no candidate

- {lane} / {category}: {what was seen and why nothing follows}

## Upstream

{filled in Step 10}

Step 8 -- Ask once. No candidates: skip to Step 10 (Upstream). Otherwise ONE AskUserQuestion with all of them, never one question per candidate:

Lane retro for {PROJECT_SLUG} on {BRANCH}: I looked at {n} of {FINISHED} finished lanes ({the lane references}) and found {k} mistakes that a machine or a link would have prevented. Details: {RETRO_FILE}.

Each proposal swaps a sentence somebody has to remember for something that works without being remembered: a check that fails when the mistake is made, or a link that saves the search.

{Numbered, one block per candidate:} {n}. {category} -- {the mistake in one plain sentence} ({lane}, {file}:{line}). Proposal: {check | pointer}: {what}. Replaces: {the line or the search}. My recommendation: {accept | skip}, because {one reason}.

RECOMMENDATION: Choose A because one reply settles all {k}, and each accepted check removes a rule from the instruction text.

A) Decide them in one reply ("1 yes, 2 no, 3 yes") -- I record each answer and plan the accepted ones Completeness: 10/10, effort: human: ~{k} min / CC: ~5 min, tier: task B) Leave them open -- the file stays, nothing is planned, the lanes are not sampled again Completeness: 4/10, effort: human: 0 / CC: 0, tier: task

Record each answer in its candidate: Decision: accepted {UTC} or Decision: skipped {UTC} -- {the human's reason, in your words}. Record only what the human answered. YTSTACK_NON_INTERACTIVE: true: do not ask and accept nothing; every decision stays open and the report lists the file under "Waiting on a person".

Step 9 -- Turn accepted candidates into tasks. Implement nothing: the instruction line a check replaces is deleted by the task that adds the check, in the same commits, not here. For each accepted candidate, add - Planned as: ... to it:

  • It fits one sentence and about three files, and the project's quick lane is on: a quick task. Write the sentence into the candidate; ytstack:quick-task takes it as it is, in a worker when a coordinator runs the plan.
  • Otherwise a task: append - [ ] T## -- {text} (lane retro {date} C#) under ## Tasks of the open, unclaimed slice plan whose files it touches, with the next free number, and raise task_count in its frontmatter. No such slice: leave it for ytstack:reassess-roadmap to place, and say so.

Step 10 -- Upstream. The candidates are the project's checks and pointers. This step is about ytstack itself: a mistake that a ytstack tool made possible or missed (the brief ytstack-brief rendered, the lane contract, a coordinator script that only polled, a missing ytstack check), or a ytstack tool that ran against a real host for the first time. The channel and its rules: ${CLAUDE_SKILL_DIR}/../../docs/ux/upstream-feedback.md; read its "The steps" and "The rules" before acting. One line per observation under ## Upstream in RETRO_FILE, with its kind and the lane evidence; none, or the project is ytstack itself: "Nothing in these lanes concerns ytstack itself." Then by UPSTREAM_FEEDBACK:

  • off: the lines stay as they are; nothing is sent or suggested.
  • suggest or unavailable: per line, search, "{UPSTREAM_TOOL}" draft, fill, check until draft ok; the line ends with suggested: <draft path> (and matches #<n> for a match). Nothing is sent.
  • file: the same, then gh issue create, or gh issue comment on the match; the line ends with filed #<n> or commented #<n>, or not sent: <the check's line>.

YTSTACK_NON_INTERACTIVE: true: nothing is sent under any setting; the drafts are listed under "Waiting on a person".

Step 11 -- Report + return.

Lane retro: {n} of {FINISHED} finished lanes sampled, {k} candidates ({a} accepted, {s} skipped, {o} open). File: {RETRO_FILE}. Planned: {the task lines and quick-task sentences, or "nothing"}. Upstream: {one line per observation: the link, or the title and draft path with "file it at Yesterday-AI/ytstack?", or "none"}.

Terminal State

Return control after the report: to the coordinator's loop when ytstack:coordinate invoked this skill, otherwise to the user. Exit without a file at the HARD-GATE.

Do NOT invoke ytstack:quick-task, ytstack:plan-task or ytstack:test-driven-development from here, and do not edit AGENTS.md, a brief template or a CI file: this skill plans, others implement. The ONLY skill you may name as a next step is ytstack:reassess-roadmap (for a task that no open slice takes).

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.