june.kim

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.

Run humanize → tighten → readability → flow → flavor → codex → sharpen in a loop until convergence. The mechanical pass before line editing.

Document shape and pacing — section titles, emoji H1s, paragraph and section sizing, bold vs italic, opening orientation. Sentence-level rhythm and cohesion live in /flow; the word layer in /humanize.

A blind-spot checker for arguments. Finds where a piece doesn't earn a sophisticated reader — across novelty, depth, rhetoric — applies the obvious subtractive fixes (reversibly), and elicits the rest via Socratic loop. Never authors new substance.

Map a foreign system onto the Natural Framework's six roles. Identifies which roles are present, which are broken or missing, and substantiates each claim against source code or documentation.

Turn a spec into a clean PR with no human intervention. Runs Volley (sharpen spec) → Merge (implement) → Hunt (verify) → Volley (clean PR) as a single pipeline.

Read sources on a foreign system, translate claims into Natural Framework terminology, and stream each into Synopsis for filing. Fan-out on thin sources. Elicit human judgment on ambiguous mappings.

Run a hypothesis graph on an engineered system. Perturb, classify the evidence trajectory, follow the edge. Fan-out for breadth, codex for adversarial filtering, e-value classification for shape. Produces a hypothesis graph document with provenance.

Look up the Parts Bin for algorithms that fill diagnosed broken roles. Produces a prescription document mapping each broken role to candidate algorithms with implementation sketches.

Create PRs from dripped entries. Human-gated. Only runs when invoked or when sweep is not in dry mode.

Run the full diagnostic pipeline on a foreign system. Intake → Diagnose → Prescribe, with mandatory human gates between each step. Skipping a step is not allowed.

Behavioral gate split out of /qa. Runs fail-on-master/pass-on-fix, routes the card by verdict, and uses the kanban trail to decide 1st-fail (→ investigate) vs 2nd-fail (→ human). Hides the test gate from qa so adversarial pride can't tangle with "did the fix work?"

Read-only design phase for feature-request (PRD-shaped) tasks (DeepSWE / Harbor). Produces the design doc whose acceptance criteria become implement-spec's proxy gate. No grader during the run (test.patch is hidden until grade time), so a missed criterion is a behavior nothing tests until too late — atomize exhaustively. No edits, no external access.

Pace PR creation — one PR open at a time per repo, one per org at a time, address feedback before pushing new. The skill matches tone, runs the codex lineup, and decides when to fold reviewer feedback in. Queue state, hard gates, gh I/O, and the heartbeat live in `sweep` CLI.

When an investigation surfaced a *separate* bug worth reporting (not the issue under investigation, but one discovered along the way), draft a fresh GitHub issue for it. SKIP when no separate bug is in the artifact. The operator approves before posting.

Implementation phase for feature-request (PRD-shaped) tasks (DeepSWE / Harbor). Builds the feature from the design doc, runs the proxy gate build-tools authored, and lets the adversary model (default Flash; configurable via $DSR_ADVERSARY_MODEL) challenge the diff against the acceptance criteria. Features are large (median ~840 LOC / 6 files) — completeness over minimalism. Proxy-green is the stopping signal available, NOT a certified grade-pass.

Read GitHub for live PR state, classify each open authored PR, and deliver one message per PR to the matching downstream actor's inbox (qa, investigate, drip, retro). OTP-style routing — pr-state is the dispatcher.

Adversarial code review on a single fix branch — codex (structural) and gemini (logic) volley until convergence or budget. The harness picks the entry, runs the test attestation, and records the verdict; the skill only does the LLM-shaped review.

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.

Build a hypothesis graph of a repo's review culture. Induces gates (instant-reject checks), signals (quality gradient), and tiebreaker (maintainer taste) from PR history. Gate before /triage.

Sift through candidates surfaced by /scout — fetch issues, screen for receipts, deposit survivors into /triage's inbox. Dedupes against ~/.sweep/seen/issues.txt to avoid resurfacing. Renamed from /prospect (and earlier /actionable).

Fan out /triage across multiple repos in parallel. One triage per repo, shared cross-repo findings, unified drip queues.

Score one issue and decide what to do with it — investigate / drop / surface / defer. Per-issue, not per-repo. Single decision in seconds; no fan-out, no nested investigations. The output feeds the investigate inbox.

Verification phase for feature-request (PRD-shaped) tasks (DeepSWE / Harbor). Runs the proxy gate (the acceptance-criteria tests build-tools authored) + the project's existing suite (regressions vs `$BASELINE_FAILS`), classifies against the design doc's acceptance criteria, writes the decision as an artifact, and emits a verdict + re-entry route. RESOLVED here means proxy-green and regression-clean, explicitly NOT a certified grade-time pass. No edits.

Compress each paragraph without losing argument. Cut filler, merge redundant sentences, shorten.

Send a file to codex (GPT-5.5) for review. Reports feedback — no changes until the user approves.

Send independent adversaries on a bug hunt — a structural pass then a logic-tracing pass, from different model families, iterated to convergence (fix, re-hunt, repeat until zero new bugs). Falls back to a Sonnet adversary when external models are unavailable or contaminated.

Scan a blog post for unlinked pop culture refs, proper nouns, named theories, and historical figures. Research and apply hyperlinks automatically.

Eliminate em-dashes (—), double en-dashes (––), and double hyphens (--) from prose. Deterministic grep for detection, subagent picks replacement punctuation and applies edits, reports reversal candidates for user review. Sister skill to /not-but.

Design review against the grimoire. Lints UI code, screenshots, or URLs for accessibility, state completeness, spacing, color, transitions, conveyance, and platform conventions.

Break a blog post into beats, check the arc, flag problems, suggest fixes. Don't change anything until the user approves.

Rewrite lazy hedges as bold narrow claims. Converges under repeated application.

Overlay a face-cam on a content video with auto-generated captions (shorts/reels format). Use when user wants to combine face video + content video with captions.

Append a timestamped entry to the project work log. Rotates at 1000 lines.

File a claim into the cached synopsis document in the subject's own vocabulary. Dedupes against existing entries, files by the subject's own structure, and maintains the S (Subjective) section. Called by Intake per-claim, not as a standalone pipeline stage.