Deliverable lint

<p align="center"><img src="https://raw.githubusercontent.com/scdenney/open-science-skills/main/plugin/skills/deliverable-open/assets/pipeline.svg" alt="The deliverable pipeline: eight steps over three layers (one wiki per piece of research, one manifest per deliverable, a shared commons)" width="900"></p>

Full guide: docs/deliverable-pipeline.md in the repository.

Detection is parallel; revision is serial. This skill finds and writes only under checks/; source files and the wiki stay untouched. The author (or sci-edit, one section at a time under its guard) revises afterwards.

Stage 0 and 1: prepare the run (shared script)

• Resolve the toolchain once, before any path below. $DTOOL is the deliverable toolchain shipped with this library at plugin/tools/deliverable. Take the first that exists: $DELIVERABLE_TOOLS; the newest installed plugin copy, $(ls -d ~/.claude/plugins/cache/open-science-skills/oss/*/ 2>/dev/null | sort -V | tail -1)tools/deliverable; the plugin/tools/deliverable directory of an open-science-skills checkout; then a legacy project_hygiene/scripts checkout. Scripts are at $DTOOL/scripts/ and kind profiles at $DTOOL/kinds/; keep them siblings, because lint_prepare.py resolves profiles relative to its own directory. If none exists, stop and name the paths you tried.

• The nearest deliverable.yml upward from cwd. Refuse to run without one; point at deliverable-open.

• python3 $DTOOL/scripts/lint_prepare.py --root <dir> --budget <usd> --mode manual|headless [--sections a,b] [--skip-fast] where $DTOOL resolves as above. It reserves checks/YYYY-MM-DD[-N]/ (never reuses a directory), runs the fast gate into <run>/fast.json (--build unless headless; --skip-fast records a skipped gate, which can never become "passed"), freezes the input hash and start time in <run>/inputs/run.json, writes <run>/inputs/index.json with every manifest section and an absolute output path per section, and one <run>/inputs/section-<id>.md per selected section containing everything a finder needs: the manifest's audience, claim, and done-test; the planning index and brief; the gate's citation status for its lines; the voice rules (resolved from the installed sci-edit skill or the writing-toolkit clone, NOT CHECKED if absent); the neighbours; the line-numbered section text; speaker notes for a deck; the knowledge-base source paths when the manifest names one; the kind profile's lint questions ($DTOOL/kinds/<kind>.md); and the exact JSON output contract.

• Read <run>/fast.json. Every FAIL becomes a P0 verbatim. If anything failed, run the adjudicator now and stop: an LLM review of a deliverable that does not build or cites a source nobody can find is spend without a return. --continue overrides.

Stage 2: section finders, in parallel

One worker per selected section, all launched together and all awaited before stage 3 (Claude: Agent calls in one message; Codex: bounded child workers queued to the slot limit). The three base questions are rule-matching against material the input file supplies (voice rules, the gate's citation status, neighbouring text), so finders run on model: "haiku", effort: "medium". When the manifest names a knowledge_base, the input adds the support question, a judgment about what a source says, so those finders run on model: "opus", effort: "low". Each worker gets one instruction: read its inputs/section-<id>.md and write the single JSON object the file specifies to the path the file names. The file carries the three questions (clarity and voice; claim-to-source; consistency) and, when the manifest names a knowledge_base, a fourth (support: does the cited source's Markdown say it; contradicted = P0 with the passage as verification). A worker sets a coverage value to false for any question it could not answer; the adjudicator counts a section as checked only when every required dimension is true. A worker that returns prose instead of the object is rerun once, then its section is NOT CHECKED.

Stage 3: the global pass

One worker (Claude: model: "opus", effort: "high"; Codex: gpt-6-astra at high, as a fresh or limited-history fork), over the assembled deliverable in section order with the same argument map. Four questions: order and flow (does the sequence carry the claim; where does a section assume what a later one establishes; for a talk, do the timings fit the slot); spec conformance (is the manifest's claim made, is the done-test met, is the register right for the audience); repetition and contradiction across sections; release rules. It writes one JSON object {"coverage": {"order":..,"spec":..,"repetition":..,"release":..}, "findings": [...]} to <run>/finders/global.json with the same finding fields; a section-local defect it notices goes in as dimension: global-local.

Stage 4: adjudicate

python3 $DTOOL/scripts/lint_adjudicate.py --root <dir> --run <run> --models finders=<model>,global=<model>,adjudicator=<this session's actual model> does the mechanical part identically for both vendors: verifies every anchor (file, line, verbatim quote; unanchored findings are dropped and counted), dedupes across dimensions, downgrades a claim-source P0 with no gate failure at that file and line and a support or consistency P0 with no verification passage, keeps the frozen start hash and marks the run stale if the inputs changed, records coverage honestly, and derives the verdict from all four conditions: fast gate passed, every section and the global pass checked, not stale, zero open P0. Then you read report.md as the domain expert: a disputed P0 needs an independent evidence check (a second lookup, a source passage; on Claude the orchestrate skill's codex-peer.sh --mode cross-check is one such check when installed); without one it stays disputed and blocking. Two models agreeing is not evidence, and a failed lookup means unresolved, never fabricated.

Budget

--budget is in US dollars, default 3; the figures are planning allowances, not prices, sized to Haiku at $0.10/$0.50 and Opus at $4/$20 per million input/output tokens: 0.01 per section on Haiku, 0.15 per section on Opus when the support question applies, 0.80 for the global pass, 0.40 for adjudication. They match lint_adjudicate.py, and a subscription session cannot enforce a hard cap. Before stage 2, if the allowance exceeds the budget, drop finders from the end of the section list and mark those sections NOT CHECKED; if the global pass alone exceeds it, stop after the gate and record every section and the global pass as NOT CHECKED. meta.cost.reported_usd is the harness's metered figure when one exists and null otherwise, never a guess.

Outputs

<run>/findings.json (meta with the frozen and current hashes, stale, models, cost, and a verdict block; coverage with the fast-gate state, whether the build was checked, every section's status and reason, the global pass, and the unanchored count; findings ranked P0/P1/P2 with id, severity, dimension, rule, file, section, line, end_line, quote, finding, action, verification, disposition, source) and <run>/report.md (verdict in the first line, the summary block, findings grouped by severity with file:line, the NOT CHECKED list, one paragraph for the author). Shipping needs the verdict's four conditions; check_deliverable.py ship refuses otherwise unless --no-lint waives the review and the receipt records the waiver.

Headless

--headless (what night-shift.py passes): no questions, no edits outside checks/; lint_prepare.py --mode headless runs the gate without --build and records the build as NOT CHECKED (shipping still builds); write both files, print the report's first ten lines, exit. If the fast gate fails, that is still a complete run.

Notes

• sci-edit manuscript is the serial, guarded path for fixes; never put the whole deliverable into one rewrite prompt.
• paper-review-lite fans out over review dimensions and presubmit runs the adversarial pipeline; both are for a manuscript at a submission milestone. This skill is the cheaper, section-anchored pass for any deliverable at any milestone, and it feeds them: run it first.
• The author reacts by dictating into inbox/ and running deliverable-intake; accepted findings become tasks, rejected ones go under Rejected findings in the declared wiki's 08-open-questions.md so they are not raised again.