Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 19 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,25 @@ All notable changes to the claude-plugins project will be documented in this fil

The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Entries are listed newest-first; each plugin section is treated as released when merged to `main`.

### vibe v1.1.94

#### Added
- New `skills/vibe/scripts/capture-guard.mjs` refuses a screenshot capture that repeats a ticket's main baseline or is too big (ISS-14158). It only ever refuses; nothing requires a capture.
- `begin --spec <spec.json>` refuses (exit 79) a main commit already captured for the ticket, any second main baseline for it, more than 240 shots (pages times states, declared in the spec), or more than about 20 minutes at the higher of 5 s per shot and the slowest rate the ticket's finished captures measured. A main commit is an ancestor of the capture worktree's origin default branch (`refs/remotes/origin/HEAD`, else `origin/main`); a ref it cannot resolve is refused.
- `end` records a running entry once as ok or failed, with its measured seconds, and refuses an entry that is not running. A running entry holds its place while its owner lives or until its own estimated minutes, never fewer than 20, have passed. `status` prints the ticket's main baseline, finished captures and measured rate. `facts` reads a ref on the host that holds the worktree.
- Overrides: the spec's `override` carries only `quote`. The rulings file comes only from `begin --rulings`. A Ruling line (read with `rulings-check.mjs`'s `recordRulings`) overrides a refusal only in a fixed shape, `- Ruling YYYY-MM-DD (<source>): "<the operator's exact words>" <TICKET> capture-guard override: <kind>`: the decision text begins with the line's only double-quoted span (two ASCII double quotes, no curly double quotes), whose content equals the spec's quote, and the ticket (exact slug, `rulingNamesScope`) and the `capture-guard override: baseline` or `capture-guard override: budget` marker follow its closing quote. Any other shape counts for nothing. `references/state.md` states the same shape for recorders. `baseline` lifts only the repeated and second main baseline refusals and `budget` only the shot and minute refusals; a capture that meets both needs both markers. A matching line without the needed marker gets its own refusal naming the marker.
- The ledger is `<ticket>.json` in `--ledger`; without it, the program's `captures` directory when exactly one fleet program's `board.json` (found through FLEET_HOME and the fleet chat bindings) lists this clone as a writer worktree, else `<git common dir>/vibe-captures`. With both `--ledger` and `--repo` the two ledgers mirror each other.
- `fleet.sh <orchDir> local|remote slot <lock> <owner> <dir> --capture <spec.json> <command...>` runs the guard against `<orchDir>/captures` (mirrored in the local clone's ledger) with `--rulings <orchDir>/rulings.md` before the load check and before any slot is taken. It times the command inside the slot, so the recorded seconds exclude the queue wait, and records how it ended. A TERM, INT or HUP stops the command and frees the slot, locally and on the remote host (a stop marker covers a remote run that has not started), and the entry records the real exit. `capture-guard.mjs` ships to the remote host. A slot without `--capture` is unchanged.
- `fleet.test.mjs` covers a `--capture` spec that is missing or unreadable (exit 2 with the usage line or `capture-guard: invalid spec`, before the load check and before any slot or command, with nothing recorded) and `--capture` given before or after `--containerized-e2e-build` (both honored, the disk preflight still runs).

#### Changed
- `references/quality-loop-visual-qa.md` gains the block `when-to-capture` ("When to capture screenshots") with ISS-14158's approved criterion 1 text: one visual-qa reviewer on the diff plus the stories by default; captures only for a shared component's pixels, a criterion or ruling that asks for visual proof, or a finding the reviewer cannot settle; one main baseline per ticket; recapture only the states a fix reaches; the final check is one visual-qa pass scoped to the changes since the last reviewed capture. It also says to run `begin` before and `end` after any capture of more than one state. `quality-loop.md` and the fleet QA gate point at it.
- The fleet QA gate's and the verify worker's before and after baseline is the ticket's one main baseline, never a new one because main moved; a difference on a state the ticket's diff does not reach comes from main and is not reported as the ticket's.
- The verify worker captures a screen when the session has a reference image for it, or when the `when-to-capture` block calls for one with `begin` and `end` around it; its shared-component sweep compares against the ticket's one main baseline. Handoff's rerun after visual QA fixes runs visual QA as one pass scoped to the changes since the last reviewed capture. The shared-scope and card-layout sentences in handoff and the verify worker now cite the visual QA reference instead of restating it.
- Fleet's Machines and CPU section adds the `--capture` slot rule, including that only an operator ruling recorded in the program rulings with the marker for the refusal it lifts lets a refused capture run; `references/state.md` adds that the marker goes on a ruling only when the operator's words ask for that override. The reporting rule and the Remaining work report template require every ETA for queued work to be computed from the remaining queue count and restated the moment a step adds more than 30 minutes.
- The README version line and the Python and vibe-bench release-pin tests move to 1.1.94; the prompt-split inventory includes the two new stubs.
- Across all plugin Markdown, including eval prompts and graders, as measured by `plugins/vibe/skills/fleet-orchestrator/scripts/instruction-words.mjs`, vibe is about 496 words longer than in 1.1.93 (origin/main at this update, c6514594).

### vibe-core v1.0.12

#### Fixed
Expand Down
2 changes: 1 addition & 1 deletion plugins/vibe/.claude-plugin/plugin.json
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
{
"name": "vibe",
"description": "Vibe sessions, handoff, seed refresh, and ClosedLoop ticket workflows",
"version": "1.1.93",
"version": "1.1.94",
"dependencies": ["closedloop-core", "vibe-core"],
"author": {"name": "ClosedLoop", "email": "support@closedloop.ai"}
}
2 changes: 1 addition & 1 deletion plugins/vibe/.codex-plugin/plugin.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "vibe",
"version": "1.1.93",
"version": "1.1.94",
"description": "Vibe sessions, handoff, seed refresh, and ClosedLoop ticket workflows",
"author": {"name": "ClosedLoop", "email": "support@closedloop.ai"},
"homepage": "https://github.com/closedloop-ai/claude-plugins",
Expand Down
2 changes: 1 addition & 1 deletion plugins/vibe/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -376,7 +376,7 @@ Run the matching Codex suite from the repository root with

### Runtime files

Version `1.1.93` contains 17 skills and 17 agents, and no hook manifest: its
Version `1.1.94` contains 17 skills and 17 agents, and no hook manifest: its
hooks, and every module they run, ship in the `vibe-core` plugin, which vibe
declares as a dependency. vibe loads those modules from vibe-core through
`shared/vibe-core.mjs`, and the commands among them keep thin wrappers at
Expand Down
24 changes: 15 additions & 9 deletions plugins/vibe/agents/vibe-verify-worker.md
Original file line number Diff line number Diff line change
Expand Up @@ -470,17 +470,23 @@ session touched:
runs displayless under the supported displayless harness. Never pass
`--allow-write`, and never use `eval` or `eval-main`.
3. For each screen: `pnpm control goto <id>`, `pnpm control wait --settled`,
`pnpm control capture --name <id>-after`, `pnpm control console --level error`,
`pnpm control capture --name <id>-after` when the session has a reference
image for that screen, or when
`../skills/vibe/references/quality-loop-visual-qa.md`, "When to capture
screenshots", calls for one, after `capture-guard.mjs begin` admits it and
with `capture-guard.mjs end` after it,
`pnpm control console --level error`,
`pnpm control console --level pageerror` and `pnpm control network`. Add
`pnpm control api GET <path>` for the record the screen shows, and
`pnpm control ipc <namespace.method>` read calls for a Desktop screen,
where relevant. For shared scope, compare each consumer page against that
page's approved prototype route; report every control present in the
prototype but missing in production, and report every control with no
source on that surface.
4. When a baseline exists (an earlier control capture of the same entry on
the merge-base or current main, never a commit that is not an ancestor),
return its artifact paths beside the new ones and say what changed. When the session recorded reference images
where relevant. For shared scope (`../skills/vibe/references/quality-loop.md`,
"Visual QA during the session", "What it checks"), report every control
with no source on that surface.
4. When a baseline exists (the ticket's one main baseline from
`../skills/vibe/scripts/capture-guard.mjs status`, never a commit that is
not an ancestor), return its artifact paths beside the new ones and say
what changed; a difference on a state the ticket's diff does not reach
comes from main and is not the ticket's. When the session recorded reference images
(`referenceImages`), open each as an image (`Read` in Claude Code,
`view_image` in Codex) beside the matching capture and report every
visible difference in shadow, elevation, font size, weight, color,
Expand All @@ -491,7 +497,7 @@ session touched:

For a shared component, capture every surface it reaches with the flag off
at 390, 412, 1024 and 1280 px wide, rendering the exact shipped story or route
and main's render of the same one; a difference is a finding.
and the ticket's one main baseline of the same one; a difference is a finding.

A static or generated page control cannot drive (an HTML page a script or a
main-process module writes, such as Desktop's sign-in return page) is still
Expand Down
3 changes: 2 additions & 1 deletion plugins/vibe/skills/fleet-orchestrator/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -154,6 +154,7 @@ Run [codex-fleet-monitor](../codex-fleet-monitor/SKILL.md) (`vibe:codex-fleet-mo

- Run heavy checks (full suites, typecheck, repo-wide lint and other gates, component story runs) through `scripts/fleet.sh <orchDir> local|remote slot heavy <owner> <worktree> <command...>`, in the worker's own worktree on that machine ([checklist step 4](./references/setup.md#checklist) gets the branch there). It caps runs at `HEAVY_SLOTS` or `REMOTE_HEAVY_SLOTS` and refuses a machine over `MAX_LOAD_PER_CORE` (exit 4): send the run to the other machine. Only the operator may override that routing by setting `FLEET_FORCE_HERE=1` with `FLEET_FORCE_REASON` to the reason they chose; the reason is written into the slot holder record. Writers never set `FLEET_FORCE_HERE`, its reason, or `FLEET_LOAD`; production admission always reads the target machine. A force is still refused above the fixed hard ceiling of four times the core count. When the force or its reason is absent, normal routing applies. Exit 78 means the containerized E2E disk preflight refused the command or the remote host did not have the exact required vibe version; its stderr names which cause. Keep the operator's daily machine for edits and git; e2e and desktop app runs never go there.
- A verified `e2e/run-containerized.sh` or `dagger call e-2-e` build receives the shared Docker backing-disk check inside the acquired slot, immediately before its command. Pass `--containerized-e2e-build` after the worktree path when a wrapper obscures that exact command shape. The check requires 64 GB decimal free; `--no-build`, lint, typecheck, tests and Storybook do not take this Docker disk gate. The remote slot verifies the installed vibe version and invokes its same admission subcommand on the owning host.
- A screenshot capture through a slot passes `--capture <spec.json>` (`ticket`, `ref`, `pages`, `states`, optional `override`) after the worktree path; `scripts/fleet.sh` runs `../vibe/scripts/capture-guard.mjs` first and refuses it (exit 79) as `../vibe/references/quality-loop-visual-qa.md`, block `when-to-capture`, says; only an operator ruling recorded for that ticket in the program rulings, carrying `capture-guard override: baseline` or `capture-guard override: budget` for the refusal it lifts and quoted in the spec's `override`, lets it run. Before approving any capture, including after a compaction, read `../vibe/scripts/capture-guard.mjs status --ticket <slug> --ledger <orchDir>/captures`.
- Before reporting or waiting on remaining work, identify true dependencies. Start independent repository selected integration, unit, typecheck and lint lanes at once through available heavy slots on allowed machines, even while e2e is queued or running elsewhere. A lane waiting only for a full slot or load moves to another allowed machine with capacity; report the move. Respect operator exclusions, machine load and slot limits, no dagger restrictions and the daily machine's no visible e2e rule. Keep the repository lane selector and `qa-impact.mjs` decision on which checks are needed.
- A shared e2e stack with fixed ports runs only under vibe's machine-wide e2e lock: `node ../vibe/scripts/e2e-lock.mjs run --head <sha> --source-hash <treeSha256> --owner <writer> --ticket <slug> <command...>` on the machine that runs it, never a second lock; receipts name HEAD plus diff/tree hashes, including dirty source. Each run gets its own compose project, so a teardown removes only its own stack; name a stack another tool started with `--foreign-compose <project>`. Unit tests, typecheck, story runs and stand-in server tests run outside it.
- `scripts/slot-lock.sh --status <lockName>` on a machine lists slot holders.
Expand Down Expand Up @@ -208,7 +209,7 @@ Hard rule: a ticket, PR body, comment, reply or other artifact for people never
## Reporting to the operator

- Surface every problem the moment you see it: limits, stalls, failures, new human review comments, blockers. Keep status updates short.
- For a remaining-work answer, compute time to ready from the dependency graph's longest dependent path, not a sum of independent lanes. Use the [Remaining work report](./references/templates.md#remaining-work-report). The response is incomplete until it states the duration and names the dependent steps and assumptions that produce it.
- For a remaining-work answer, compute time to ready from the dependency graph's longest dependent path, not a sum of independent lanes. Use the [Remaining work report](./references/templates.md#remaining-work-report). The response is incomplete until it states the duration and names the dependent steps and assumptions that produce it. Every ETA for queued work is computed from the remaining queue: state how many runs are still queued or running and each one's estimate, and restate the ETA the moment a step adds more than 30 minutes.
- Before reporting a count or an "only" claim from a command you ran, check that it exited 0 and that no `head`, `tail` or tool truncation cut its output; otherwise rerun the narrow command or call the figure partial. Principle 2 holds for your own findings too.
- When the operator steps away (vibe's `away-protocol` hook puts this checklist in context on "brb", "stepping away" and the like, and the return checklist on "I'm back"), keep `<orchDir>/attention.md` current (decision needed, recommendation, what is blocked), keep reading idle writers' last messages, and lead with that list on return.
- Put every finding the operator reviews on one page using the stable `D`, `X`, `N`, `S` and `R` IDs in the [operator review template](./references/templates.md#operator-review-page); never restart numbering across audits.
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -2,11 +2,14 @@

Moved out of `../SKILL.md`, which says when to read it; later edits to that text land here. Markdown links are relative to this file; other paths stay relative to the skill folder.

> Read `../vibe/references/quality-loop-visual-qa.md` (block `when-to-capture`) before any screenshot capture or visual QA pass, in the loop or at handoff.

<!-- block: qa-gate-rules -->
- QA happens before merge; never call a PR ready to enqueue before its QA is done.
- A unit's PR is ready only once its review of the completed work passed (`references/ticket-units.md`, "Review the completed work"): every member and PRD criterion with a code row and a capture for each page and surface it names.
- Each PR moves through independent review, fixes, verification, visual review where needed, one feature-complete QA pass, operator QA, one main merge, then enqueue.
- QA evidence records screen, host, state and flag arm. Later diffs use handoff's `qa-impact.mjs` with importers, FEATURE_MAP and lanes. Obey its scope; `None` runs no local lane, screenshot, audit or critic. CI is per push. This also governs the final main merge before `shepherd.mjs`.

- Off by default and an open owner decision: with `FLEET_REVIEW_HANDOFF_ONLY=1`, `node scripts/review-chain.mjs --pr-head <sha> --handoff-reviewed-head <sha the writer's handoff review reports covering>` prints `verify-fixes-only` when that review covered the PR's exact head, and the orchestrator then runs only the verification of the fixes in place of its own independent review. Any other output, an unset variable or a head that moved after the handoff review keeps the full chain above.
- Before an audit, reconcile operator chat and AskUserQuestion answers plus guide-session and QA records into `<orchDir>/rulings.md`, then run `node ../handoff/scripts/rulings-check.mjs --program <orchDir>/rulings.md --record <transcript-or-record> ...`; audit the acceptance criteria together with those rulings.
- Before showing `<orchDir>/review.md`, run `node scripts/review-page-check.mjs <orchDir>/review.md`.
Expand All @@ -26,5 +29,5 @@ Moved out of `../SKILL.md`, which says when to read it; later edits to that text
- When a repository's Storybook CI requires a release-version bump and history entry for new stories, the writer includes them in the reviewed pre-PR candidate so that lane can pass. Pin main once before the final handoff checks as vibe:handoff describes; after QA, settle a version-only conflict once at enqueue in merge order and rerun the affected lane. Do not repeatedly merge main merely to chase another open PR's reserved version.
- A fix reaches a preview only once the ref that preview deploys from contains it: check `git merge-base --is-ancestor <fix> <ref>` before calling it live.
- Track merge order for stacked or overlapping PRs and tell each writer which one merges first.
- A before and after comparison uses the candidate's merge-base or current main as the baseline, never an unrelated branch.
- A before and after comparison uses the ticket's one main baseline (`../vibe/scripts/capture-guard.mjs status`), never an unrelated branch and never a new baseline because main moved.
<!-- end: qa-gate-rules -->
Loading
Loading