task-pipeline-skill 1.96.0 → 1.98.0

package/CHANGELOG.md

@@ -1,3 +1,93 @@
+## v1.98.0 (2026-10-20)
+
+Guards: 448 → **448**; property checks 15 → **15**.
+
+A cheaper run that stops looping, and the operators' retro insights in the retro doctrine.
+Measured on 2026-10-20: before any code a run was told to read about 236 KB of doctrine
+(the command 14 KB restating `SKILL.md` 21 KB, plus companion-skills, model-tiering,
+continuity, stages, stage 0, grill, knowledge-sources and gates.md "at every stage"); a
+small fix met eleven stages because the short path was only proposed and silence took the
+full flow; every node closed through three blind tiers, R2 owed two rounds, the
+certification ceiling only measured, the review loop ran to five rounds; and the retro
+floor stage 0 reads in full ran up to ~76k tokens in one project, 27 620 bytes here.
+
+**A. Proportional run.** Stage 0's short-path triage now decides instead of proposing:
+when the gap check leaves no material unknown and the change is R1, at most five files,
+no public contract, no user-facing change and no deploy, the run takes the short path and
+says so in one line (brief in the run ledger, build with tests, one verification, docs
+and handoff, a retro entry only for a new lesson). `scripts/short_path.py` computes it.
+The example's gates at stages 0, 2, 3, 7 and 10 are `auto`; the operator's answer or the
+standing authorization recorded in the brief is the go, and a project that wants the wait
+sets a gate `manual` in its own `pipeline.json`.
+
+**B. Reading diet.** At entry a run reads `SKILL.md` only (its stage table is the index);
+each stage's section of `stages.md` and its detail load when the stage starts; grill,
+knowledge-sources, companion-skills, continuity, model-tiering and gates.md load on named
+triggers. The command is a 1.5 KB pointer. Measured, the floor before code on the short
+path is 35 KB (was 236 KB; 80 KB when the grill and knowledge-sources triggers fire);
+`SKILL.md` is 16.2 KB, 3 753 cl100k tokens of body (was 4 709, cap 5 000). The three descriptions are 386, 399 and 395 characters; every trigger phrase
+they dropped moved verbatim into a `## Triggers` section of the same `SKILL.md`.
+
+**C. Anti-loop.** `graph.py certify`: an R1 node certifies with one blind tier; R2 owes one
+passing round, R3 two (one blind, one adversarial), a flagship surface two with one
+adversarial; a round of the same kind on an unchanged tree is refused; at the ceiling the
+node is parked with its evidence (exit 7) instead of measured. `waves.py finish
+--clean-rounds` (default 1; 2 for an R3 or flagship iteration). The fix loop and the
+review loop cap at three rounds and stop.
+
+**D. Retro doctrine.** The floor (standing instructions plus run stamps) has a 12 000-byte
+budget, a standing row at most 600 characters, a stamp at most 300; stage 0 prints the
+bytes and bytes/4 tokens, gate 10 fails over budget. `scripts/retro_budget.py`
+(`check`, `migrate --write`, `verify`) moves narrative to the archive verbatim with a
+sha256 per block; this repository's floor went from 27 620 to 6 532 bytes that way. Each
+entry names its Class (13 classes) and Cost; a standing row is searched against
+`learned.md` and the lesson index first; a lesson standing in two projects goes upstream
+in the same run. `learned.md` gains rules 33 to 39 from the 2026-10-20 audit of ~54
+retrospectives. evidence-docs' 34.7 KB copy of the retro doctrine is now a pointer.
+
+Tests: `graph_test.py` 171 → 173 cases (tier count by risk, unchanged-input refusal,
+ceiling stops), `waves_test.py` 135 cases with `--clean-rounds`,
+`test/audit_regressions/diet_retro_budget.py` (9 cases, over-budget plants must fail) and
+`diet_proportional.py` (6 cases). Twelve negative plants retargeted to the new wording
+(stage 0 and 10 gates name their authorization and the hand-back instead of `manual`).
+
+## v1.97.1 — 2026-10-17
+
+The same source as v1.97.0, plus its release declaration. The v1.97.0 tag failed its own release run
+([38976060549](https://github.com/ssheleg/task-pipeline/actions/runs/38976060549)): `retro.md` named no
+release after its newest stamp, so the validator on the tag's tree refused and `publish` was skipped. Tags
+are immutable here, so v1.97.0 stays as a failed tag and this patch carries the declaration (the v1.94.0 →
+v1.94.1 precedent). No doctrine change.
+
+## v1.97.0 — 2026-10-16
+
+Source candidate; not published.
+
+Guards: 447 → **448**; property checks 15 → **15**.
+
+The look now runs on WebKit first. Every browser example in the bundle opened the
+tool's default engine, a Chromium-family browser, and `chrome-devtools` drives Chrome
+only — so the look the pipeline taught was a Chromium look reported as *tested in a
+browser*, and a layout that breaks only in Safari passed it. `references/browser.md`
+opens with *Engine order — WebKit (Safari) first*: Safari through `safaridriver` for a
+real browser, Playwright's `webkit` for automation (`playwright-cli open --browser
+webkit`, the MCP's `--browser webkit`, read from each tool's `--help`), Safari Web
+Inspector for debugging, and `open -a Safari` instead of the system default browser.
+Chromium is a stated exception — Lighthouse, a CDP trace, a heap snapshot, a
+Chrome-only API — named with its reason, on a temporary profile the run closes; a
+surface checked only in Chromium is reported as exactly that. The recipe, the
+sign-in example and the library example (`const { webkit } = require('playwright')`)
+now open WebKit.
+
+`references/companion-skills.md` makes `playwright` on WebKit the default look and
+`chrome-devtools` the stated Chromium exception, in the matrix, the preflight lines
+and the one rule for both channels; stages 5, 6 and 8, `tdd.md`, the README, the
+Cursor rule, the skill card and the contact-sheet template say the same. `DEC-0007`
+records the order: the engine is ranked, the tools' quality still is not, so
+`DEC-0004` keeps its other two parts. The new guard holds the section above the model
+and the recipe's `--browser webkit`; its negative self-test strips the flag and
+requires the validator to refuse.
+
 ## v1.96.0 — 2026-10-15
 
 Source candidate; not published.

package/HOW-IT-WORKS.md

@@ -28,11 +28,12 @@ you asked for.
 ## 1. How a request finds the pipeline
 
 An agent choosing skills sees exactly one thing: the `description` field in
-`SKILL.md`. It is capped at **1024 characters** and currently uses **1008** — that
-headroom is the budget every future addition pays from, which is why it is
-tracked as a board row rather than left to be discovered.
+`SKILL.md`. It is capped at **1024 characters**, and since 1.98.0 it is the short
+form (`B-001` on the board tracks the measured length): every host preloads it into
+every session. The full trigger list, unchanged, sits under `## Triggers` in the same
+file, where the family's routing hook and this repository's checks read it.
 
-It is written in three parts, and the middle one was learned the hard way:
+That list is written in three parts, and the middle one was learned the hard way:
 
 - **Work that changes the repository** — a feature, fix, refactor, migration,
   integration, rewrite, adoption or hardening, in English and Russian both.
@@ -125,23 +126,26 @@ it has been forgotten.
 | | Stage | What must be true to leave it | Gate |
 |---|---|---|---|
 | 1 | Docs study | every contract the design will lock is grounded on **fetched** docs, not recall | auto |
-| 2 | Brainstorm + decompose | design approved, every REQ answered by it. A platform is cut into modules, **walking skeleton first**, every REQ in exactly one module | manual |
-| 3 | Spec | contracts locked. User-facing work runs three tracks — what it **does**, how it **sounds**, how it **looks**. A declined track is recorded, never silent | manual |
+| 2 | Brainstorm + decompose | design approved, every REQ answered by it. A platform is cut into modules, **walking skeleton first**, every REQ in exactly one module | auto |
+| 3 | Spec | contracts locked. User-facing work runs three tracks — what it **does**, how it **sounds**, how it **looks**. A declined track is recorded, never silent | auto |
 | 4 | Plan | the REQ set-comparison holds: brief REQs == union of `Implements:` | auto |
 | 5 | Build | TDD per task, a review after each, findings fixed or parked **with a ruling** | auto |
 | 6 | Tests | the **full** suite green; a web surface checked in a browser, not in the diff | auto |
-| 7 | Lint + deploy | the authorization is specific, and the CI verdict is **read** before any tag | manual |
+| 7 | Lint + deploy | the authorization is specific, and the CI verdict is **read** before any tag | auto |
 | 8 | Post-deploy | the verification trio, not one of three; a verification row per shipped REQ | auto |
 | 9 | Docs + wiki | **three** artifacts — module docs, the wiki, and the code graph | auto |
-| 10 | Acceptance | the ladder walk first, then the table, then the retrospective | manual |
+| 10 | Acceptance | the ladder walk first, then the table, then the retrospective | auto |
 
-### Why the manual gates are where they are
+### Why every gate ships `auto`
 
 `auto` means the agent verifies the check itself and proceeds. `manual` means it
-waits for an explicit go. The four manual gates sit at the four places where a
-wrong turn is expensive and hard to reverse: **what we are building** (2), **the
-contracts** (3), **the outside world** (7), and **whether it is done** (10).
-Everywhere else, stopping to ask costs more than it saves.
+waits for an explicit go. Until 1.98.0 the example made stages 0, 2, 3, 7 and 10
+manual, the places where a wrong turn is expensive: **the brief**, **what we are
+building**, **the contracts**, **the outside world** and **whether it is done**. Under an
+operator's standing "decide and act" instruction each of them was a pause nobody
+answered, so the example ships them `auto`: the approval they name is read from the
+operator's answer or the standing authorization recorded in the brief, and a project
+that wants the wait sets the gate `manual` in its own `pipeline.json`.
 
 ---
 

package/README.md

@@ -182,7 +182,7 @@ and the inline fallback satisfy the same [provider contract](plugins/task-pipeli
 | 3 Spec | [`spec.md`](plugins/task-pipeline/skills/task-pipeline/references/spec.md) — UX-track order, locked contracts, global constraints, self-review |
 | 4 Plan | [`planning.md`](plugins/task-pipeline/skills/task-pipeline/references/planning.md) — zero-context tasks, parallel groups, no placeholders |
 | the queue | [`work-graph.md`](plugins/task-pipeline/skills/task-pipeline/references/work-graph.md) — a script walks the graph so the model never reads it: 400 nodes and 4 print the same 27-byte frontier |
-| closing a node | [`certification.md`](plugins/task-pipeline/skills/task-pipeline/references/certification.md) — three blind readings at escalating visibility (the changed code, what reaches it, the product around it), all three required to pass — and a fourth, `verifier-visual`, on a flagship or product surface, which reads the contact sheet against the director record; a failing round records itself and the node stays open |
+| closing a node | [`certification.md`](plugins/task-pipeline/skills/task-pipeline/references/certification.md) — three blind readings at escalating visibility (the changed code, what reaches it, the product around it), all three required to pass (one on an R1 node; a clean round ends a node unless it is R3 or flagship) — and a fourth, `verifier-visual`, on a flagship or product surface, which reads the contact sheet against the director record; a failing round records itself and the node stays open, and at the ceiling the node is parked with its evidence |
 | 5 Build | [`build.md`](plugins/task-pipeline/skills/task-pipeline/references/build.md) + [`review.md`](plugins/task-pipeline/skills/task-pipeline/references/review.md) — isolation, ledger, subagent loop, review rubric, fix loop |
 | 5–6 TDD | [`tdd.md`](plugins/task-pipeline/skills/task-pipeline/references/tdd.md) — the iron law, red/green/refactor, the suite gate |
 | 5, 6, 8 The browser | [`browser.md`](plugins/task-pipeline/skills/task-pipeline/references/browser.md) — the ref model both channels share, the four commands the look is made of, sessions, and the three different things *"tested in a browser"* means |
@@ -221,9 +221,11 @@ Before any technical work, task-pipeline closes every question the codebase and
 project's records can answer, then asks you the rest in **one batch**, ordered by
 dependency, each with a recommended answer — one at a time only where the answers
 depend on each other or you ask for it — until every decision branch is resolved and
-locked into a **task brief**. There is
-no "clear enough task" exemption: no stage-1 work starts without a committed,
-confirmed brief.
+locked into a **task brief**. The depth is proportional: a small change (R1, at
+most five files, no public contract, no user-facing change, no deploy) takes the
+**short path** by default, printed in one line: the brief in the run ledger, build
+with tests, one verification, docs and handoff. Anything larger starts stage 1 only
+on a committed, confirmed brief.
 
 **Domain awareness.** While exploring, the grill reads the project's own
 `CONTEXT.md` / `docs/adr/` and holds you to them — calling out terms that conflict
@@ -280,14 +282,17 @@ and neither can a **browser test suite**, which asserts what someone wrote down
 counted as coverage rather than as the look. Absent, the run says *"verified by reading
 the diff"* — a weaker claim, recorded as one.
 
-There are **two channels and either is enough**, ranked by nothing: `playwright`
-(`npm install -D @playwright/cli@latest`, or
-`claude mcp add playwright npx @playwright/mcp@latest`) needs no plugin and is already
-there on a project that runs Playwright, and its CLI half puts no tool schema in the
-context window; `chrome-devtools`
-(`/plugin install chrome-devtools-mcp@claude-plugins-official`) is the one that reaches
-past the page to Lighthouse and heap snapshots, and analyses a performance trace rather
-than only recording one. Neither is a gate,
+There are **two channels and either is enough**, and the **engine** is ranked:
+**WebKit (Safari) first, Chromium a stated exception**. `playwright`
+(`npm install -D @playwright/cli@latest` then `playwright-cli open --browser webkit`, or
+`claude mcp add playwright -- npx @playwright/mcp@latest --browser webkit`) is the default
+look: it reaches WebKit, needs no plugin, is already there on a project that runs
+Playwright, and its CLI half puts no tool schema in the context window; on macOS Safari
+itself is driven through `safaridriver`. `chrome-devtools`
+(`/plugin install chrome-devtools-mcp@claude-plugins-official`) drives Chrome only and is
+reached for when the question needs Chrome — Lighthouse, heap snapshots, a performance
+trace analysed rather than only recorded — said with the reason; a look taken only there
+is reported as *checked in Chromium*. Neither is a gate,
 and a CLI, a library or a backend service is never offered either.
 
 **The wiki is [obsidian-wiki](https://github.com/ar9av/obsidian-wiki)** (Karpathy's
@@ -479,9 +484,9 @@ iteration runs in this order:
 - the survey;
 - module dossiers written in parallel, laid onto one canvas until a pass changes nothing;
 - questions asked in rounds, never one at a time, while agents keep working;
-- a parallel build, each node checked the moment it lands: one round of three blind
-  readings for every node, and two or three passing rounds by its risk class (R2, R3 — one
-  of the three recorded adversarial);
+- a parallel build, each node checked the moment it lands: one clean round ends it (one
+  blind reading on R1, three on R2), and an R3 or flagship node owes a second, adversarial
+  round;
 - one release, checked as released;
 - the next pool, handed to you grouped by module.
 
@@ -503,7 +508,8 @@ every repeat pass logs one line per touched file with **the reason that forced i
 
 It trips on revert-oscillation, a file edited twice for the same reason, a
 resurrected finding, a third entry into one stage, or two loops editing one file —
-plus hard caps (5 fix rounds per task, 2 re-entries per stage, 3 passes per module).
+plus hard caps that stop (3 fix rounds per task, 3 review rounds, 3 failing certification
+rounds per node, 2 re-entries per stage, 3 passes per module).
 On a trip the run **stops editing**, names shape A and shape B with their evidence,
 escalates to the layer that owns the conflict (rubric → operator → plan → spec →
 module map), re-plans the check as an ordered checklist with one verification
@@ -635,7 +641,7 @@ must say so.
 ### Held to Anthropic's own Skill authoring guidance
 
 Audited against the four Agent Skills pages. Most of it already held — `name`
-13/64 chars, `description` 887/1024 chars, `SKILL.md` 275/500 lines, all 53 references
+13/64 chars, `description` 386/1024 chars, `SKILL.md` 207/500 lines, all 53 references
 reachable from `SKILL.md` through explicit reference routes, and the bundle far under the 30 MB ceiling. What
 did not, now does:
 

package/SKILL-CARD.md

@@ -12,9 +12,9 @@ harmless.
 |---|---|
 | **Purpose** | Runs a substantial task through ten gated delivery stages — intake grill, docs study, brainstorm, spec, plan, subagent build, tests, lint/deploy, post-deploy, docs+registers, acceptance — refusing to advance until each gate passes |
 | **Owner** | ssheleg ([github.com/ssheleg/task-pipeline](https://github.com/ssheleg/task-pipeline)) |
-| **Version** | 1.96.0 |
+| **Version** | 1.98.0 |
 | **Surface** | Claude Code (filesystem skill + plugin) and the vercel `skills` CLI. **Not** uploaded to the Skills API; custom Skills do not sync across surfaces |
-| **Dependencies** | None required. Optional: `context7` (MCP), `figma` (MCP), super-ux, agent-sync, graphify, obsidian-wiki, and **one of two browser channels** — `playwright` (CLI or MCP) or `chrome-devtools` (MCP); either satisfies the browser step and neither is required. Every stage's doctrine ships in-repo; user-facing stage 3 requires a valid scenario artifact; super-ux is preferred, with equivalent-provider and inline-fallback paths under the provider contract |
+| **Dependencies** | None required. Optional: `context7` (MCP), `figma` (MCP), super-ux, agent-sync, graphify, obsidian-wiki, and **one of two browser channels** — `playwright` (CLI or MCP) or `chrome-devtools` (MCP); either satisfies the browser step and neither is required. The engine is WebKit (Safari) first; `chrome-devtools` is Chrome-only, a stated exception. Every stage's doctrine ships in-repo; user-facing stage 3 requires a valid scenario artifact; super-ux is preferred, with equivalent-provider and inline-fallback paths under the provider contract |
 | **Evaluation status** | Suite authored, 5 categories. One recorded run, **self-observed by the author**; **zero blind runs on zero of three models** — the split, and the numbers, live in [`evals/RESULTS.md`](evals/RESULTS.md) and are computed by `evals/run.py` |
 
 ## Risk-tier disclosure

package/cursor/rules/task-pipeline.mdc

@@ -67,9 +67,14 @@ event, not the end of the run — what loses work is crossing it mid-item.
 
 ## Stage 0 — intake grill (MANDATORY, do this FIRST)
 
-Never skipped, and nothing to install — the grill is part of this rule. No "the
-task was already clear" exemption, no starting stage 1 while the user thinks. A
-one-line task ("build feature X") is not enough to finish autonomously.
+Never skipped, and nothing to install: the grill is part of this rule. Its depth is
+proportional. **A small change takes the short path by default** when the gap check
+leaves no material unknown and the change is R1 risk, at most five files, no public
+contract, no user-facing change and no deploy: say so in one line, write the brief and
+its REQ rows into the run ledger, build with tests, run one verification, update the
+docs and handoff, and write a retro entry only for a new lesson. Anything larger runs
+the full flow, and a one-line task ("build feature X") is not enough to finish it
+autonomously.
 
 **Phase 1 — harvest the sources BEFORE the first question.** Find what the project
 already knows about this task and read it: the code; **the code graph if one is
@@ -98,7 +103,8 @@ like any other source, and it **points while the code decides** — never the
 tiebreaker.
 **Read `docs/evidence/retro.md` in full if it exists** — its *standing
 instructions* are what previous runs of this pipeline got wrong in this project, and
-they bind this run (they are capped at ten, so reading them all is cheap). Stamp
+they bind this run (ten rows, ten stamps, and a 12 000-byte budget for the two, so
+reading them all is cheap; say the byte count). Stamp
 each one the moment it fires: that date is the only evidence behind the
 cold-retirement rule at stage 10.
 If no wiki is installed, recommend it once and continue:
@@ -199,7 +205,7 @@ not authorize an outward, irreversible action — stage 7 stops and asks.
 ## Stages 1→10
 
 1. **Docs study** (auto) — ground every external lib/API on current docs, not recall.
-2. **Brainstorm + decompose** (manual) — read the brief, explore the code, then 2–3
+2. **Brainstorm + decompose** (auto; the brief's standing authorization is the go) — read the brief, explore the code, then 2–3
    approaches with a recommendation, YAGNI applied, design presented in sections and
    approved. **Hard gate: no code, no scaffolding before that approval** — including
    on "obviously simple" tasks. Record whether the task touches a user-facing surface
@@ -222,11 +228,12 @@ not authorize an outward, irreversible action — stage 7 stops and asks.
    re-measure, re-cut — and 9–10 close the release. No wave builds a shape a later
    wave undoes. **Around the waves runs the release iteration** ("release iteration" /
    "итерация релиза"): module dossiers integrated into one canvas, questions in batches,
-   each node verified as it lands in one to three rounds by its risk class (R1–R3), a
+   each node verified as it lands, one clean round ending it unless it is R3 or a
+   flagship surface (R1 on one blind tier; the ceiling parks a churning node), a
    wave's stage 7 integrated without tagging, one release gated and verified as
    released, the next pool handed back; a finding made mid-iteration joins the
    inventory rather than being fixed alone.
-3. **Spec** (manual) — write and commit the design; lock every shared contract
+3. **Spec** (auto; the brief's standing authorization is the go) — write and commit the design; lock every shared contract
    (types, schemas, signatures, file layout) plus a **Global Constraints** block of
    exact values every later task inherits. Self-review it for placeholders,
    contradictions, ambiguity and scope before showing it. For UI tasks, run the UX
@@ -272,15 +279,19 @@ not authorize an outward, irreversible action — stage 7 stops and asks.
    look. With a browser channel connected: load the surface, snapshot it, read the
    console and the network log, and quote what you read. Fix what the look finds
    before calling the stage green. Without one, say "verified by reading the diff" and
-   treat that as the weaker claim it is. Two channels, either is enough, ranked by
-   nothing: `npm install -D @playwright/cli@latest` (or
-   `claude mcp add playwright npx @playwright/mcp@latest`) needs no plugin, and its CLI
-   half puts no tool schema in the context window;
-   `/plugin install chrome-devtools-mcp@claude-plugins-official` is the one that
-   reaches Lighthouse and heap snapshots, and analyses a trace rather than only
-   recording one. Never a gate. A CLI, a
+   treat that as the weaker claim it is. Two channels, either is enough, on WebKit
+   (Safari) first with Chromium a stated exception: `npm install -D @playwright/cli@latest`
+   then `playwright-cli open --browser webkit <url>` (or
+   `claude mcp add playwright -- npx @playwright/mcp@latest --browser webkit`) is the
+   default look, needs no plugin, and its CLI half puts no tool schema in the context
+   window; on macOS, Safari itself through `safaridriver`, and `open -a Safari <url>`
+   rather than the system default browser for a person.
+   `/plugin install chrome-devtools-mcp@claude-plugins-official` drives Chrome only and
+   is reached for when the question needs Chrome — Lighthouse, heap snapshots, a trace
+   analysed rather than only recorded — said with the reason; a look taken only there
+   is reported as "checked in Chromium", never "tested in a browser". Never a gate. A CLI, a
    library or a backend service is never offered either.
-7. **Lint + deploy** (manual) — lint clean AND suite green before deploy; deploy is
+7. **Lint + deploy** (auto; the brief's standing authorization is the go) — lint clean AND suite green before deploy; deploy is
    outward → explicit user go, or the specific standing authorization recorded in
    the stage-0 brief.
 8. **Post-deploy** (auto) — tail logs / health-check; clean boot or an honest
@@ -316,7 +327,7 @@ not authorize an outward, irreversible action — stage 7 stops and asks.
     propose the edit and get an explicit go, or carry it over with the exact change
     written down. A doc that was worth reading at stage 0 and is wrong now is the
     next run's false premise.
-10. **Acceptance** (manual) — the closing stage, in two halves.
+10. **Acceptance** (auto; the brief's standing authorization is the go) — the closing stage, in two halves.
     **First the ladder walk**, because the REQ table only finds what was named and
     lost: a comparison needs two sides and **an absence has one**. Walk each REQ
     bottom-up through its rungs — recorded decision → spec section → contract *and
@@ -414,9 +425,10 @@ not authorize an outward, irreversible action — stage 7 stops and asks.
     of the next run reads those standing instructions **in full**, which is why the
     cap is not negotiable and the prune is a gate item, not a good intention.
     Gate: ladder walk ran, no REQ `unknown`, no ledger row without
-    a home, every repository clean/pushed/pointed-at, user signs off, **and the
-    retro is written — stamped first, then pruned before anything was added, at or
-    under its cap, every deletion logged, and the counts printed beside this verdict**.
+    a home, every repository clean/pushed/pointed-at, the hand-back written, **and the
+    retro is written: stamped first, then pruned before anything was added, at or
+    under its cap and its byte budget, every deletion logged, and the counts printed
+    beside this verdict**.
 
 Cross-cutting: answer from the brief's autonomy section rather than re-asking, log every deferral in the ledger, never narrow the task silently, track
 tasks, conventional commits, honest degradation (never claim a failed/skipped step
@@ -432,8 +444,8 @@ file, the pass, and the reason that forced it — a finding id or a failed gate
 "Cleanup" is not a reason. **Stop immediately** when: an edit undoes what an earlier
 pass in this run deliberately did (A→B→A); the same file is edited twice for the
 same reason; a finding already fixed or parked comes back; a stage is entered a
-third time for one artifact; two loops edit one file. Hard caps: 5 fix rounds per
-task, 2 re-entries per stage, 3 passes per module. On a stop: name shape A and shape
+third time for one artifact; two loops edit one file. Hard caps, each a stop: 3 fix
+rounds per task, 2 re-entries per stage, 3 passes per module. On a stop: name shape A and shape
 B with their evidence and who asks for each; escalate to the layer that owns the
 conflict (two findings → the review rubric; finding vs plan → the user; plan vs spec
 → back to planning; ambiguous spec → back to spec; two modules claiming one file →

package/evals/task-pipeline.evals.json

@@ -476,9 +476,9 @@
       ],
       "query": "wave 1 of a release iteration has two nodes: N-001 (R1: fix two wrong links in docs/README.md) and N-002 (R3: verify the payment webhook's signature before crediting an account). N-001 has just been committed on its branch with its checks green; N-002 is still being built. Continue.",
       "expected_behavior": [
-        "Starts N-001's certification now, without waiting for N-002 or the end of the wave: one round — the three blind tiers plus one fix cycle — and N-001 is closable after that round passes",
-        "Keeps N-002's risk class at R3 (money, a security check) and never lowers it to save time; it owes three passing rounds, one of them recorded adversarial (`certify --kind adversarial`), planting a defect and trying abuse cases, on another model where one is available",
-        "Runs every owed round in full even when round 2 comes back clean — the adversarial round asks a different question and always runs",
+        "Starts N-001's certification now, without waiting for N-002 or the end of the wave: one round on one blind tier (R1) plus one fix cycle, and N-001 is closable after that round passes",
+        "Keeps N-002's risk class at R3 (money, a security check) and never lowers it to save time; it owes two passing rounds, one blind and one recorded adversarial (`certify --kind adversarial`), planting a defect and trying abuse cases, on another model where one is available",
+        "Runs the adversarial round even when the blind round comes back clean, because it asks a different question, and never re-runs a blind round on the unchanged tree",
         "Treats `graph.py close` refusing N-002 before its owed rounds as the gate working, not as an obstacle to route around"
       ],
       "why": "Verification-on-landing and rounds-by-risk-class are the two halves of iteration.md → Verification starts when a node lands. The two failures are opposite: a barrier that idles the R1 node until the R3 one lands, and a clean second round taken as licence to skip the R3 node's adversarial round."

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "task-pipeline-skill",
-  "version": "1.96.0",
+  "version": "1.98.0",
   "description": "Full-cycle delivery pipeline for coding agents: a mandatory built-in intake grill, then 10 gated stages (docs, brainstorm+decompose, spec, plan, build, tests, lint/deploy, post-deploy, docs/wiki, acceptance). Every stage's doctrine ships inside the skill — no companion plugin required. This package is the installer CLI.",
   "bin": {
     "task-pipeline": "bin/task-pipeline.js"

package/plugins/task-pipeline/.claude-plugin/plugin.json

@@ -2,8 +2,8 @@
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "task-pipeline",
   "displayName": "Task Pipeline",
-  "description": "Runs a substantial task through a mandatory built-in intake grill, then 10 gated stages (docs, brainstorm+decompose, spec, plan, subagent build, tests, lint/deploy, post-deploy, docs/wiki, acceptance). Every stage's doctrine is built into the skill — no companion plugin required — with typed auto/judgment/manual gates, a frozen requirement spine that closes with evidence, a work board and a verification ledger that outlive a run, an exposure line naming what shipped unconfirmed, a progress rail computed from the project's own config, a loop guard whose review ceiling measures rather than stops, and stage-3 tracks for what a product does, how it sounds and how it looks. Two modes need no task: `checkup` (what is unverified) and `setup` (audit existing docs); a third, `waves`, takes the whole board and closes it in waves that each finish in one pass. Retro insights can publish upstream as issues, opt-in and redacted.",
-  "version": "1.96.0",
+  "description": "Runs a substantial task through a mandatory built-in intake grill, then 10 gated stages (docs, brainstorm+decompose, spec, plan, subagent build, tests, lint/deploy, post-deploy, docs/wiki, acceptance). Every stage's doctrine is built into the skill — no companion plugin required — with typed auto/judgment/manual gates, a frozen requirement spine that closes with evidence, a work board and a verification ledger that outlive a run, an exposure line naming what shipped unconfirmed, a progress rail computed from the project's own config, a proportional short path for small changes, a loop guard whose ceilings stop and record a blocker, and stage-3 tracks for what a product does, how it sounds and how it looks. Two modes need no task: `checkup` (what is unverified) and `setup` (audit existing docs); a third, `waves`, takes the whole board and closes it in waves that each finish in one pass. Retro insights can publish upstream as issues, opt-in and redacted.",
+  "version": "1.98.0",
   "author": {
     "name": "ssheleg",
     "url": "https://x.com/sshlg93"