メニュー
The pipeline's backward pass. Read what happened, compress repeated episodes into durable artifacts (skill patches, parameter files, memories). Lossy by design — the losses are what didn't generalize. Naming the loss is part of the artifact.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Filter recent Hacker News for live threads on June's topics and hand him the list. Separately, flag the rare subset where work he has already done settles a specific quoted claim, and draft that comment as a plain argument with no link to his own writing. Use when June asks to check HN or find threads worth reading or commenting on.
Scan blog posts for AI writing tics and missed opportunities for human voice. Reports subtractions (AI patterns to remove) and additions (wordplay, arc, and claims to strengthen).
Poll June's Hacker News feeds, report what is new since the last run, and idempotently arm a background watch for replies to his comments. State persists in ~/Documents/sweep/ so it survives across sessions. Use when June asks what is new on HN, whether anyone replied, or to start watching a thread.
Argument-level compression by a strengthened absence test (does the argument stand with the same force, clarity, evidence, scope, pacing, and misreading-resistance?). Removes passages that fail it (self-referential cruft, cross-section restatement, self-recap, re-derivation, scaffolding, expired reader-orientation, stakes inflation, inert hedges, redundant examples); grafts good rhetoric onto an adjacent essential point rather than deleting. Idempotent. The structural cousin of /tighten (word-level) and /not-but (negation).
Audit a technical candidate's resume/CV by verifying its claims against the public contribution graph instead of taking the text at face value. Sorts every claim onto the self/peer/world attestation lattice, applies a cost-and-stake filter so cheap proxies (stars, downloads, backlinks) don't count, runs LIVE checks (GitHub, arXiv, package registries, CVEs), and reports two separate axes — merit (staked acceptance of the candidate's specific work) and attention (timestamped output) — plus a verifiability score in [0,1]. For technical recruiters and hiring managers. Invoke when given a candidate — a GitHub handle, a resume/CV, an arXiv author, a personal site — and asked to verify, audit, vet, or check their claims. Runs at two depths — a quick triage scan or a deep per-claim audit — and elicits which. Checks the provenance of a personal site or self-hosted resume (Internet Archive + git history) as an un-backdatable anti-tailoring anchor. Needs tool access (Bash/curl or WebFetch).
Find and defuse "not X but Y" / "isn't X; it's Y" / "X — not Y" constructions. Deterministic grep for detection (no LLM blind spot), the not-but subagent for triage. Split-or-cut tail negations ("X, not Y") aggressively — promote the point to its own sentence or delete it; recast in place only mid-clause "not X but Y" that genuinely can't be split; keep just the rare iconic thesis line literal.
| name | retro |
| description | The pipeline's backward pass. Read what happened, compress repeated episodes into durable artifacts (skill patches, parameter files, memories). Lossy by design — the losses are what didn't generalize. Naming the loss is part of the artifact. |
| argument-hint | <repo> [--since YYYY-MM-DD] |
| allowed-tools | Read, Write, Edit, Bash, Glob, Grep |
You are the consolidation pipe running over the cycle that just finished. Your job: read the logs, find the repeated episodes, compile each repetition into the smallest artifact that lets the next cycle skip that work. Then say what you dropped.
Same morphism every time:
watch yourself (or the agents) repeat
→ compile the repetition into a skill / param / memory / CLI
→ next cycle runs
→ notice where the compiled artifact failed
→ update
See Memory Compression. The pattern is recursive: this skill is itself a compression of the manual retros that came before it.
| What you need | How to get it |
|---|---|
| All structured events for a repo since date | sweep retro gather --repo … --since YYYY-MM-DD |
| Pipeline policy commits (split events by version) | sweep retro policy-changes --since YYYY-MM-DD |
| Existing parameter file for a repo | sweep retro params --repo … |
| PR outcomes (own + prior art) from GitHub | sweep retro outcomes --repo … |
| Append a parameter update | sweep retro set --repo … --key … --value … --reason … |
Record a fix_ready entry | sweep retro fix-ready --repo … --issue N --line-count N --description … |
| Skill-use stats (which skills fired, how often, with what outcome) | sweep retro skill-stats --since YYYY-MM-DD |
| Propose an eviction (skill / memory not earning its keep) | sweep retro evict --target <name> --reason … |
| List CLI commands agents tried to call but that don't exist | sweep retro missing-calls --since YYYY-MM-DD |
NOT BUILT — mid-run-edits is spec'd below but unimplemented; don't reach for it | |
NOT BUILT — ref-drift is spec'd below but unimplemented; don't reach for it | |
NOT BUILT — em-dash-audit is spec'd below but unimplemented; don't reach for it |
The three rows above are planned, not built.
sweep retro --helpis the source of truth for what exists. The prose below specifies their intended behavior (for whoever builds them); until then, skip steps that depend on them and don't burn calls reaching for them.
Before averaging anything across the window, run sweep retro policy-changes --since <window-start>. Each row is a commit that touched classifier rules, routing, model defaults, workflow code, or a skill spec — anything that changes what the pipeline does with the same input.
Use those timestamps to slice the event stream into eras: events before commit <ts> ran under the old policy, events after ran under the new. Mixing them collapses two different systems into one number and you'll prescribe a fix for an average that no run actually saw. When you can't separate cleanly, say so — name the policy change in S, attribute the events to "pre-X" and "post-X" in O, and let A treat them as separate populations.
If a CLI command isn't there yet, that's a gap — don't reimplement the gh / jsonl plumbing here. The deterministic harness rejects malformed inputs with errors you close the loop on.
Each compression has four parts. Write them all.
The repeated episode. Name what kept happening. "Three triage runs in a row killed good first issue items because fast claimers raced us." If you can't name three episodes, it's not a pattern — log it as ambiguous, don't compress yet.
The compiled artifact. Where the lesson now lives. Pick the lightest form that binds the next cycle:
| Compression target | Where it lives | Binds when |
|---|---|---|
| Single repeated action → operation | Skill markdown patch (/triage, /drip, etc.) | Skill is invoked |
| Repeated parameter setting | sweep retro set JSONL | Skill reads params |
| Cross-session insight | Auto-memory entry | Next conversation loads memory |
| Repeated deterministic check | CLI subcommand or gate | Every invocation |
| Repeated skill sequence | Composed skill (level 2) | Sequence invoked |
Prefer the lightest. A param file beats a skill patch beats a new CLI command beats a new skill. Move heavier only when the lighter form keeps decompressing back into work.
The loss. Name what you discarded. "Three specific reviewer quotes; the exact PR numbers; the time-of-day they were rejected. Kept: kill-list addition for good first issue on repos with >100 stars." The loss is what didn't generalize. If you can't name the loss, you're not compressing — you're filing.
The level. Tag which level of the tower this compression lives at (see Memory Compression):
Most retros produce L0→L1 patches. Watch for L1→L2 opportunities (you keep running A then B then C — make /abc). L2→L3 is rare and the work you do least often.
sweep retro gather returns structured events from every pipeline skill plus PR outcomes (own + prior art). You don't grep log files; the CLI assembles. If you need an event type that isn't there, add it to the logging contract in the relevant skill, don't reach around the CLI.
Special inputs that don't fit the gather shape:
sweep retro missing-calls returns every sweep <subcmd> … invocation an agent tried that failed with "command not found" or "unknown subcommand," grouped by the requested call and ranked by frequency. Each missing call is a wishlist entry the agent demonstrated by reaching for it. This is how the CLI grows: not from spec, from agent demand. Three agents on three repos all tried sweep drip checkup --repo X → build it next. The skill markdown stays honest because failures are visible.sweep retro mid-run-edits surfaces ad-hoc changes agents made to skill specs during the cycle. Each one is a compression opportunity — the agent already noticed the pattern; retro's job is to formalize it (and remove the ad-hoc patch, or fold it in deliberately).action.yml description + any text the action emits as PR comments, blog posts cross-referenced from outbound. Specifically watched: pr-to-merged, internal-reasoning-of-prose-compiler, speedrunning-open-source, the-natural-framework — these are linked from issue templates and READMEs, so a rename or stale section breaks the outbound chain. NOT outbound: skills/*.md, hypothesis/retro graphs, worklogs, ~/.sweep/*, internal bootstrap/roadmap docs, CLAUDE.md. The distinction is whether a maintainer evaluating us reads it, not whether it's checked into a public repo.sweep retro ref-drift walks the outbound surface and verifies every https://github.com/... path and #anchor still resolves. A renamed heading or moved file silently turns a load-bearing pointer into a 404 the maintainer sees first. The check fetches each unique target once per pass, caches for 24h, and reports broken refs grouped by source file. Triage these as obvious compressions. A broken outbound link is a receipt against our own pipeline. Fix the link or the heading; don't ship another issue against a 404.sweep retro em-dash-audit greps the outbound surface for em-dashes (—) and double-hyphens (--) in prose, excluding code blocks, table cells, and reference-list separators. Budget is zero. This is a Van Halen rider: the PR Quality Gate flags em-dashes as a low-effort signal, so an outbound artifact containing one proves we don't dogfood. The maintainer doesn't read it as voice; they read it as a competence canary. Internal artifacts (skills, retros, memory) are out of scope; em-dashes that face us are fine.~/.sweep/repos/<owner>-<repo>/RETRO_GRAPH.md). One graph per repo, one entry per PR, classifying which meta-hypotheses the outcome supports or refutes. This is a long-running compression artifact, updated each pass.Fan-out compresses human judgment: things three independent perspectives converge on are obvious; things they diverge on are ambiguous. Retro's compression gate is the same. The bucket a finding lands in determines whether the artifact gets written this pass.
Obvious (fold immediately):
sweep retro fix-readysweep retro setAmbiguous (stash for human):
Ambiguous items append to the worklog tagged [AMBIGUOUS] with one-line tension descriptions. Each ambiguous→obvious transition over time is real progress. The retro pipeline gets faster when the ambiguous bucket shrinks, not when more memories get written.
Compression has a counterpart: artifacts that don't earn their keep get pruned. Without this, the pipeline accumulates dead weight and slows.
Each pass, run sweep retro skill-stats --since 30d. For each pipeline skill, parameter, and memory:
sweep retro evict --target <name>. Human approves; CLI removes.Guard rails (rules that prevent bad outcomes even when they never fire) are protected — cooldown_until, "fail on master, pass with fix," "fork-only push." Annotate them as guards so eviction skips them.
The metric retro optimizes for is time × quality preserved per artifact. High product → keep / promote. Low product → evict. Same metric for skills, params, memories.
sweep retro gather --repo <r> --since <d> — pull all events.sweep retro outcomes --repo <r> — pull PR outcomes (own + prior art) since last pass.sweep retro missing-calls --since <d> — pull the CLI-wishlist (commands agents reached for that don't exist). Triage these first — every entry is the strongest possible evidence for a CLI gap: an agent already needed it. Rank by frequency, name the verb-noun shape, queue as CLI work.
3a. (when built) sweep retro ref-drift and sweep retro em-dash-audit — walk the outbound surface, report broken refs and any em-dashes in prose. Both block the ship, not the retro. A broken link or a stray em-dash in an outbound artifact is a receipt-failure against our own pipeline. These two are not yet implemented — skip this step until they appear in sweep retro --help.sweep retro … CLI.RETRO_GRAPH.md for the repo.sweep retro skill-stats + evict proposals.[AMBIGUOUS].cooldown_until is the one parameter retro never auto-resets; humans clear it.action.yml, cross-linked blog posts) is a competence canary. Broken refs, em-dashes in prose, and other low-effort tells fail the canary before the maintainer reaches the argument. Outbound-surface checks block the ship, not the retro. Internal artifacts (skills, retros, memory, worklogs) are out of scope; only the surface the maintainer sees matters.sweep retro gather does, with proper schema validation.)