| name | re-frame2-pair-retro |
| description | Retrospect on a `re-frame2-pair` session and turn it into prioritised improvement ideas for the pair skill, scripts, MCP surface, or upstream `re-frame2` Tool-Pair contract; optionally drafts a GitHub issue the user can file. Activates on two triggers: (a) **explicit pull** — user asks for a retrospective on a recent pair session ("retro on this pair session", "what went wrong with my pair session", "review my re-frame2-pair session", "draft an issue about that"); or (b) **post-error within a re-frame2-pair session** — after a stack trace, failed dispatch, red CI, or a runtime error during live pair work, to post-mortem the firefight. Requires evidence: a concrete `re-frame2-pair` session in this conversation (turns where the user attached, dispatched, walked traces/epochs, hot-swapped, or time-travelled), **or** a user-supplied recap of one. **Not** for ordinary `re-frame2-pair` operation, nor for code/spec/framework work the body's routing matrix sends elsewhere. Vocabulary matches alone ("retro", "what went wrong", "any improvements?") do not justify activation — a real pair session must have occurred or be recapped.
|
| allowed-tools | ["Read","Grep","Glob","Write","Bash(gh issue list *)","Bash(gh issue view *)","Bash(gh issue create *)","Bash(gh label list *)","mcp__re-frame2-pair__discover-app"] |
re-frame2-pair-retro
Turns a re-frame2-pair session — the one happening now (post-error), a just-finished one, or one summarised by the user as a recap — into a product retrospective for re-frame2-pair. This is a conversation, not an automated report: surface findings, let the user steer which ones matter, then converge on improvements.
Two entry modes
The two triggers enter the workflow differently:
- Explicit pull — the user asked for a retro ("retro on this session", "draft an issue about that"). Run the analysis workflow below directly.
- Post-error post-mortem — a stack trace, failed dispatch, red CI, or an
:rf.error/* event fired during live re-frame2-pair work and you are reaching for this skill unprompted. The user has not asked for a retrospective, so do not dump the full retro output mid-firefight. First confirm the fire is out — fixing the immediate runtime failure is re-frame2-pair's job (route there). Then offer the retro in one line ("Want me to retro on what made that error hard to chase?") and run the workflow only on a yes. Its subject is the workflow friction the firefight exposed (why the error was hard to find, recover, or trust), never the application bug. If the user declines, stop — a post-error trigger is an offer, not an obligation.
When you cannot tell which mode you are in, treat it as post-error: offer rather than assume.
When NOT to use this skill
Story recorder-session retros are out of scope. A retro on a Story Test Codegen recording belongs in re-frame2-pair's variant-refinement workflow (the recorder output is a :play-script snippet to refine against a frame, not a pair-session friction trace). If the user asks to "retro on my recorded play sequence" or similar, decline and route to re-frame2-pair.
Routing decisions (mid-session pair work, app-authoring without a live runtime, framework / spec feedback, app-bug help, vocabulary-only matches) follow the matrix at skills/README.md §Skill routing — single source and §Disqualifiers. The activation precondition — a real re-frame2-pair session must have occurred or be recapped — lives in the frontmatter description and that matrix.
When in doubt, ask: "Was there a re-frame2-pair session you want me to retrospect on? If you can paste a short recap I can work from that." Decline rather than fabricate evidence.
Guard rails
- Always start with session analysis. Do not jump to fixes. Surface friction points before root causes, and let the user pick which ones to dig into.
- Default to diagnosis, not contribution. Do not assume the user wants to file a GitHub issue or propose a patch. The default tool grant is read-only —
Read, Grep, Glob, plus gh issue list / gh issue view. Mutation (gh issue create) is granted but gated by the approval rule below.
- Never file a GitHub issue without explicit user approval. Drafting issue text is fine; running
gh issue create is not, until the user has seen the draft and said go. The skill carries no Edit — source rewrites in another repo are out of scope; route those as issue suggestions. Its only Write use is composing the issue body for --body-file (the shell-safety mechanics live in ../shared/issue-filing.md; the operational gate is §Filing improvements below).
- Live-runtime probes are opt-in. The default path is transcript-only; the skill does not probe the live runtime by default. The allow-list grants exactly one re-frame2-pair MCP tool — the read-only
mcp__re-frame2-pair__discover-app — use-gated, not default: reach for it only when the retro is tied to an in-conversation live session already attached AND the user has confirmed a probe. It captures build id/health/session sentinel and (with the server's tools/list) sanity-checks tool availability. Any deeper live work (dispatch, app-db read, epoch walk) is pair-programming, not a retro — route to the re-frame2-pair skill and reason from the transcript here. Recap-only/offline retros never probe.
- Stay focused on improving
re-frame2-pair. If the right fix is upstream in re-frame2 — a gap in a Tool-Pair surface from ../shared/tool-pair-surfaces.md — say so, name the specific surface (not "the contract"), and route the proposal to a GitHub issue against re-frame2.
- Tracker boundary — file GitHub issues, never
bd beads. bd is the re-frame2 monorepo's internal tracker; skills consumed downstream file against the target repo's GitHub issues via gh issue create. The full filing recipe lives in ../shared/issue-filing.md; §Filing improvements below is the re-frame2-pair-retro specialisation.
- Do not propose fixes via
re-frame-10x. v2's pair tooling does not depend on it. Time-travel and trace-stream consumption ride directly on re-frame2's Tool-Pair surfaces — the canonical surface enumeration and the "supersedes re-frame-10x" claim live in ../shared/tool-pair-surfaces.md.
Working style
Diagnostic posture rules (evidence over vibes; symptom vs cause; direct/indirect friction; positive gaps; creatively ambitious after diagnosis) live in references/working-style.md. Apply them per finding.
Analysis workflow and output shape
Load ../shared/retro-protocol.md for the workflow shape — the diagnosis-first steps (read the evidence → identify friction candidates → route to the detection rule → surface findings with concrete evidence → cross-link the canonical fix → opt-in issue-filing → confident, no-hedging voice), plus the evidence-citation discipline, the untrusted-evidence and universal-redaction boundaries, the layer-routing rules, and the seven-section output shape. It is shared with re-frame2-improver; the pair-retro specialisation is below.
Session-evidence contract
The diagnosis is only as trustworthy as the evidence boundary it runs on. Conversation order is not causal order; two builds, two frames, or two attach attempts are not one retry loop. Before reconstructing the timeline, bind the evidence to one causally-ordered session:
- One evidence envelope. Scope the retro to a single session — the user-stated session or recap, or the contiguous pair workflow serving one goal. When two plausible envelopes are present (two goals, two builds, a recap alongside a live session), ask which session to review rather than merging them.
- Build a causal ledger, not a transcript-order list. Associate each result with its initiating call, and keep the provenance the evidence carries: build id, frame id, the runtime-instance / freshness token (session sentinel), and whether a fact came from native conversation turns or a user recap. Arrival order alone is never causal order — a delayed or background result belongs to the call that issued it, not to whatever ran most recently.
- Exclude unrelated activity. Background worker runs, CI results, shell commands, code-review threads, and app-authoring edits are out of scope unless the user explicitly names one as pair-session friction. Red CI from another job is not this session's regression.
- Supersession. A later successful retry or an explicit target switch supersedes the earlier state; the earlier failure survives only as friction that actually cost effort, never presented as the current / final tool state.
- Unknown over inferred. A missing, truncated, unmatched, or still-running result is
unknown/incomplete — never scored as a success or a failure. State the limitation, and ask for the missing result only when it would change a finding.
- Attribution. Keep recap claims marked as user recap; never invent turn numbers, timestamps, or tool-payload fields that were not supplied.
This sharpens the protocol's step 1 (read the evidence in scope) and its evidence-discipline rule for the session shape — the reconstruction below runs on the bounded ledger, not the raw transcript.
Pair-retro deltas:
- Reconstruct the session goal and a short timeline first — the intended outcome, environment facts (platform, target repo, live runtime state, tooling constraints), and the turns where progress stalled, restarted, detoured, or needed a workaround (tool errors, empty/stale outputs, retries, clarification loops). Present the friction as a numbered list before classifying, and ask which to dig into.
- Route each finding through the catalogues. Classify one primary root cause from the canonical taxonomy in
references/analysis-lenses.md §Root-cause categories (single source — do not redefine inline). Pattern-match recurring friction against references/known-frictions.md to tell a one-off from a product gap — including its error-observability class when the session chased an error (why it fired, where it surfaced, or why the framework's typed recovery wasn't what the user expected).
- Generate improvements at the right layer — skill wording, structured op, runtime surface, cross-platform behavior, validation/fixture, instrumentation, or an upstream
re-frame2 GitHub issue. Prefer proposals that remove repeated effort, not just this session's exact symptom.
- Output uses the protocol's seven sections under these labels:
Goal; Observed friction (numbered, first, for user steering); Likely root causes (one primary per finding, contributors allowed); Improvement ideas (2-5; default mix 1-3 grounded + 0-2 bolder, each carrying the friction it addresses, why re-frame2-pair wasn't enough, the proposed change, the layer, and a one-line impact); Bolder ideas; Issue candidates (only if the user wants them); Other possibilities. If the session is too thin, say so plainly and ask for a recap or permission to use a longer conversation as input.
Filing improvements
Filing is a two-step, approval-gated mode — distinct from the default diagnose-only mode:
- Default mode (no approval needed). Read the transcript, surface findings, draft issue text inline in the conversation. Use
gh issue list / gh issue view to check whether an existing issue already covers the friction.
- Filing mode (explicit approval required). Only after the user has seen a draft and explicitly says "file it" (or equivalent), invoke
gh issue create against the appropriate repo. Never run gh issue create on your own initiative — Bash(gh issue create *) is granted solely to enable this user-approved transition. Offer filing only if useful, and split into multiple focused issues when the findings warrant it.
Routing. Both pair-tool friction and framework friction file against day8/re-frame2 (the monorepo that ships the pair tool); carry the tool-vs-framework distinction in the title + body, naming the specific surface from ../shared/tool-pair-surfaces.md for a framework gap. The full pre-drafting routing decision is references/issue-template.md §Routing first.
Labels are optional taxonomy, not a filing precondition. gh issue create fails the whole command on an unknown --label, and the target repo may not define retro / pair-mcp / upstream-from-re-frame2-pair. Detect with gh label list, pass a --label only after confirming the repo defines it, and fall back to a no-label gh issue create so the handoff always lands. This is the canonical statement of the rule; the operational degrade steps are in references/issue-template.md §Filing with gh issue create.
Filing mechanics. The shared shell-safety core — search-before-file, the Write-tool + --body-file body path (a fresh per-filing OS-temp file), the safe-alphabet --title / --search, and redaction — lives once in ../shared/issue-filing.md; the body skeleton is references/issue-template.md.
Anti-patterns
- Don't reduce every problem to "write more docs". Consider product behavior, tooling, defaults, instrumentation first.
- Don't confuse a transient local outage with a product gap unless the workflow made recovery harder than it should have.
- Don't propose vague improvements like "better UX" without naming the concrete missing behavior.
- Don't file speculative issues unsupported by the session, or pressure the user to file anything.
