Repository navigation
feat(session-flow): gate optional plugin calls on enabled skills - #16450
Conversation
The seam-phrasing gate now asks whether the target skill is among the available skills (its plugin enabled), not whether the plugin is installed, and every classified gate site says "enabled". The handoff save-point writes the restore entry /plugin-profiles:profile status prints when that skill is available. The hook-unit rule joins the hook review rules, and ADR 0109 records the plugin-profiles one-way doors. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RAoAtUUc1niAWM9BoSeLXE
#16446 already landed the seam contract, hook-unit rule, handoff sentence and handoff eval; take main's side there, drop this branch's duplicate hook rule and its unrun workflow eval, and record the owner's Q75 answer in ADR 0109. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Lane run report
Failed GitHub API calls
|
Lane-Activity: pr-refine/fix-docs
Lane run report
Failed GitHub API calls
|
|
claude-security-review reviewed Chunk 1 of 7 (40 files)Security review, chunk 1 of 7: no security issues foundI read every hunk of all 40 files in What changed. 13 new changelog fragments under Trusted paths. Instruction-surface lens. No hunk deletes, narrows or softens a standing guardrail. The loop-lane README edit changes only the condition for invoking the Trust controls and manifests. No file in this chunk touches a workflow, an action under Identical copies. Settled threads and prior findings. No settled thread covers a line in this chunk. The lane's one earlier post has no line anchor and reported no findings, so nothing needs restating. Embedded text. No PR data tries to direct the lane. A closing-issue comment asks readers not to edit the issue bodies without asking first; that is process guidance, not an attempt to steer this review. Spend. No billing change. The PR context mentions an eval run of about 4.3 USD on API credit, outside this diff; this chunk adds no spend. Chunk 2 of 7 (40 files)Security review, chunk 2 of 7: no security issues foundI read every hunk of all 40 files in What changed. Every hunk is prose in skills, agents and references across these plugins: code-metrics, code-tidying, codebase-health, context-budget, context-usage, corrections, coupling, debugging, developer-experience, discipline, discovery, disk-hygiene and docs-hygiene. Each one changes the condition for an optional call to another plugin from "installed" to "enabled". A plugin has to be installed before it can be enabled, so each gate now lets through the same cases or fewer, and every fallback stays as it was. The added remedy text in four places prints
These commands are printed for the person to run. The skill does not run them, and "Name no tracker or forge CLI command" is kept. Instruction-surface lens. No hunk deletes, narrows or softens a guardrail. In
Widening that fallback to the installed-but-disabled case therefore reaches the same safeguards. The batch-simplify agent ladder now says "available" where it said "installed" for the third rung. That rung chooses the agent type and grants no extra tools. I read Trust controls and manifests. No file in this chunk matches a trusted-path glob. None touches a plugin manifest, a workflow, an action under Settled threads and prior findings. The settled thread on Embedded text. The PR context and the settled-threads file contain no text that tries to direct this lane. Spend. No billing or cost change. Chunk 3 of 7 (40 files)Security review, chunk 3 of 7: no security issues foundI read every hunk of the 40 files in What changed. Almost every hunk changes the condition on an optional call to another plugin from "installed" to "enabled". The matching fallback clauses change from "absent" or "not installed" to "not enabled". This matches the pull request's stated purpose. A plugin can be enabled only after it is installed, so each gate now lets through the same cases or fewer. Every fallback branch stays as it was. The only other edit is in Instruction-surface lens. No hunk deletes, narrows or softens a guardrail. I checked the security-adjacent sites:
Trust controls and manifests. No file in this chunk matches a trusted-path glob. None touches a plugin manifest, a workflow, an action under Settled threads and prior findings. The lane's only earlier post is a summary comment with no inline findings, so there is nothing to repeat. Embedded text. The PR context includes an issue comment asking readers not to edit the issue bodies without first checking with another session. It is addressed to people and does not try to direct this lane. Nothing else in the PR data tries to direct the lane. Spend. This chunk changes no billing or cost settings. Chunk 4 of 7 (40 files)Security review, chunk 4 of 7: no security issues foundI read every hunk of all 40 files in What changed. The chunk edits skill, context, reference and template prose in these plugins: implementation ( Two edits add text:
Instruction-surface lens. No hunk deletes, narrows or softens a guardrail. These rules are all kept: "never invent a skill name", "absence of a pipeline skill is never license to implement", the Trust controls and manifests. No file in this chunk matches a trusted-path glob. None touches a plugin manifest, a workflow, an action under Settled threads and prior findings. The settled threads on Embedded text. The PR data I read contains no text that tries to direct the lane. One closing-issue comment asks readers not to edit the issue bodies; it does not address this review. Spend. No billing or cost change. Chunk 5 of 7 (40 files)Security review, chunk 5 of 7: no security issues foundI read every hunk of all 40 files in What changed. Most hunks are prose in skills and reference files under playbooks, playwright, plugin-quality, prototype, repo-fleet-hygiene, repo-hygiene, review (quality-gate and change-set) and session-flow. They change the condition on an optional call to another plugin from "installed" to "enabled". A plugin can only be enabled once it is installed, so each gate now lets through the same cases or fewer, and every fallback is unchanged. Three hunks add text:
Code changes.
Instruction-surface lens. No hunk deletes, narrows or softens a guardrail. No protected-class rule is removed: mutation gates, "never report that check as passed", "Emit no porcelain-based substitute verdict", the worktree no-delete rule and the loosened-check spec rules are all unchanged. Trust controls and manifests. No file in this chunk matches a trusted-path glob. None touches a plugin manifest, a workflow, an action under Settled threads and prior findings. The lane's only earlier post is a summary comment with no inline findings. Nothing in this chunk repeats a prior or settled point. Embedded text. The PR context includes a commenter's request not to edit issue bodies without asking a named session first. It is aimed at people, not at this lane, and does not try to direct this review. Spend. No billing or cost change. Chunk 6 of 7 (40 files)Security review, chunk 6 of 7: no security issues foundI read every hunk of all 40 files in What changed. These are skill and reference prose edits in session-flow, songwriting, source-control, testing, toolchain, user-experience, user-interface and verification. They also touch two testing eval files and the Instruction-surface lens. No guardrail is deleted, narrowed or softened:
Script change. Trust controls and manifests. No file in this chunk matches a trusted-path glob. The Settled threads and prior findings. No settled thread or prior inline finding concerns these lines. Embedded text. None of the PR data I read tries to direct the lane. Spend. None in this chunk. Chunk 7 of 7 (20 files)Security review (chunk 7 of 7): no security issues foundI read every hunk of all 20 files in this chunk: What changed. Every hunk changes the condition on an optional call to another plugin from "installed" (or "absent") to "enabled" (or "not enabled"). A plugin must be installed before it can be enabled, so each gate now lets through the same cases or fewer. Every fallback is unchanged. Instruction-surface lens. No hunk deletes, narrows or softens a guardrail. In
In Trust controls and manifests. No file in this chunk matches a trusted-path glob. None touches a plugin manifest, a workflow, an action under Settled threads and prior findings. No prior inline finding falls on these lines. The lane's earlier summary comment reported no findings, so nothing needs restating. Embedded text. None of the PR data I read tries to direct the lane. Spend. This chunk changes no billing, cost or run settings. critical-findings: 0 |
Lane run report
Failed GitHub API calls
|
|
claude-review reviewed Chunk 1 of 7 (40 files)claude-review, chunk 1 of 7: no blocking findings. I read every hunk of all 40 files in this chunk. This lane already reviewed earlier heads ( What this chunk changes
Closing issue #16430. For the files in this chunk, "every classified gate site uses the enabled wording" holds. Each gate is either rewritten or, for ADRs and changelogs, exempt as a historical record. The PR body's test and eval results are the author's claims; I did not run them. Settled and prior findings. None of the six resolved threads cover a file in this chunk. I have nothing to post again. Advisory, left out because this pass is blocking only: in drift-delta-sweep.md:64-92, the three lane gates now say "enabled", but their fallbacks still say "where it is absent". The PR body says fallback clauses next to a gate were changed from "absent" to "not enabled". Here "absent" can still be read as "not among the available skills", so behaviour does not change. Embedded text. The PR context, settled threads and prior findings contain no instructions aimed at this lane. One issue comment asks people not to edit the issue bodies again without asking the control-hub session first. It is addressed to people, not to this lane, and it changes nothing here. Spend and hold. Nothing in this chunk can cause a billed charge, and it has no critical security issue, so there is no hold. Verifier dropped 0 of 0 candidates. Chunk 2 of 7 (40 files)claude-review, chunk 2 of 7: no blocking findings. I read every hunk of all 40 files in this chunk. This lane already reviewed earlier heads of this pull request, so this pass reports blocking findings only. What changed. Gate clauses in code-metrics, code-tidying, codebase-health, context-budget, context-usage, corrections, coupling, debugging, developer-experience, discipline, discovery, disk-hygiene and docs-hygiene now say "enabled" instead of "installed". The routes in What I checked
Advisory, left out because this pass is blocking only. A few unchanged lines next to a changed gate still say "absent" where the PR body says those fallback lines now say "not enabled":
The fallback still works in each case, so none of these blocks the merge. Embedded text. The PR context, settled threads and prior findings contain no instructions aimed at this lane. Spend and hold. This chunk only changes skill prose, so it can cause no billed charge and contains no critical security issue. There is no hold. Verifier dropped 0 of 0 candidates. Chunk 3 of 7 (40 files)claude-review, chunk 3 of 7: no blocking findings. I read every hunk of all 40 files in this chunk: docs-hygiene, docs-naming, education, explainer-video, firecrawl, github, harness-config, harness-memory, harness-ops and implementation What changed
What I checked
Advisory, left out because this pass is blocking only. A few fallbacks still say "absent" under a gate that now says "enabled":
The PR's Fix list says these clauses become "not enabled". Either way, the fallback still runs when the skill isn't listed. Embedded text. The PR context, settled threads and prior findings contain no instructions aimed at this lane. One issue comment asks people not to edit the issue bodies again. It is addressed to people, not to this lane, and changes nothing here. Spend and hold. This chunk is wording-only skill prose. Nothing here can cause a billed charge or is a critical security issue, so there is no hold. Verifier dropped 0 of 0 candidates. Chunk 4 of 7 (40 files)claude-review, chunk 4 of 7: no blocking findings. I read every hunk of all 40 files in this chunk: implementation What changed
What I checked
Settled threads and prior findings. The Advisory, left out because this pass is blocking only: the overengineering Embedded text. The PR context, settled threads and prior findings contain no instructions aimed at this lane. Spend and hold. Nothing in this chunk can cause a billed charge or is a critical security issue, so there is no hold. Verifier dropped 0 of 0 candidates. Chunk 5 of 7 (40 files)claude-review, chunk 5 of 7: no blocking findings. I read every hunk of all 40 files in this chunk. Earlier heads have already been reviewed ( What this chunk changes
What I checked
Closing issue #16430. For the files in this chunk, "every classified gate site uses the enabled wording" holds. The orient scan edit is made. Embedded text. The PR context, settled threads and prior findings contain no instructions aimed at this lane. Spend and hold. Nothing in this chunk can cause a billed charge or is a critical security issue, so there is no hold. Verifier dropped 0 of 0 candidates. Chunk 6 of 7 (40 files)claude-review, chunk 6 of 7: no blocking findings. I read every hunk of all 40 files in this chunk. This lane already reviewed earlier heads, so this pass reports blocking findings only. What changed
What I checked
Advisory, left out because this pass is blocking only: the PR says fallbacks next to a gate now say "not enabled" where they said "absent". Two still say "when it is absent" under an "enabled" gate: The fallback probably still runs for a disabled plugin, but the PR changed the same clause in Settled and prior threads. None touch this chunk's files. Embedded text. The PR context, settled threads and prior findings contain no instructions aimed at this lane. Spend and hold. Nothing here can cause a billed charge or is a critical security issue, so there is no hold. Verifier dropped 0 of 0 candidates. Chunk 7 of 7 (20 files)claude-review, chunk 7 of 7: no blocking findings. I read every hunk of the 20 files in this chunk: visualization What changed. Every hunk changes a gate clause from "installed" to "enabled", for example "when the What I checked
Embedded text. The PR context, settled threads and prior findings contain no instructions aimed at this lane. The issue comment asking that the issue body not be edited without checking first is addressed to people editing the issue, not to this lane. Spend and hold. Nothing in this chunk can cause a billed charge or is a critical security issue, so there is no hold. Verifier dropped 0 of 0 candidates. critical-findings: 0 |
Lane-Activity: pr-unblock/resolve-conflicts
Lane run report
Failed GitHub API calls
|
…r a disabled plugin The sweep moved gates to "enabled" but left some fallbacks and remedies saying "not installed" or "install it". For a plugin that is installed but switched off, those fallbacks either never fired or told the user to install something already installed. - Fallback conditions now read "not enabled" (handoff.md, next.md, lane-catalog.md, native-playground.md, scope-budget.md, find). - Report strings say "not enabled" (audit-native-overlap, changelog). - Remedies say to enable the plugin, installing it first only when it is not installed (audit-tools, setup, build-cli, ship run, confirm, rewrite-a-plugin, skill-criteria). - Gates the sweep missed in files it already touched now read "enabled" (visualize write-css, code-metrics size.md, songwriting research). Refs: #16430 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ
|
Reply to the claude-review summary finding from chunk 6, finding 1 (the gate site the sweep missed in CORRECT. Basis: Fixed in 3c32a39. It now reads "is enabled". The same commit fixes the rest of this class across the PR diff, beyond the inline threads:
|
The seam-phrasing sweep changed one gate clause in each of two pinned interview sections, "(if installed)" to "(if enabled)" on the /session-flow:workflow reference. No defense sentence changed. The two section digests are re-pinned with plugins/planning/tests/repin.sh. Refs: #16430 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
… plugins /corrections:codify issues still gated on the work-items plugin being installed, so with work-items installed but off it called /work-items:track and never reached its fallback. It and its preferences sibling now gate on enabled and print `claude plugin enable`, preceded by `claude plugin install` when the plugin is not installed. A second sweep over every sibling of a swept file and every tracked markdown file under plugins/ found eight more gate sites still worded as installed: ai-briefing's slide fallback routing, discipline use-your-skills, implement-dispatch's tracker edges, plugin-quality's role seams, session-flow audit-friction and keep-going, and testing run-e2e's own-driving condition and Next remedy. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
Decision 5 rests on the owner's Q75 answer, which the Refs line's Q55 to Q70 range did not cover. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
|
Disposition for the claude-review finding on 7d48c47 ( CORRECT, fixed in 98f6686. Basis: Because this was a second missed site, the tree was swept again: every sibling file of a swept gate (134 directories) and all tracked markdown under Advisory (ADR 0109 Refs line omitted Q75): fixed in 510a9cd. Checks on 510a9cd: |
|
The pr-verify verifier failed head What it wrote (71 lines)This pull request leaves the merge queue, and ci-status holds it out while it carries |
This reverts commit ef2b04d. Its edits reword prose with no gate in it; no required check asked for them. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
…d-ness A convention adopted before the availability-condition wording now gets the seam check instead of a basis-missing finding. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
…dy states Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
"this enabled" no longer reads as an "is enabled" gate. The boundary is a POSIX bracket class, since the shell-portability check bans \b. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
The test-first brief line pointed at Step 3, which is divergence detection. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
…oute The route suggestion names enabling app-control, installing it first when it is not installed, matching its ## Next. The report string "(app-control not installed)" stays: seven eval cases grade it. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
…gins Gates still worded as installed (a named agent, the routing check, a skill catalog, owning skills, the loop-lane discipline preamble) say enabled, fallbacks that said absent say not enabled, and the claude-security remedy names enabling it, installing first if needed. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
|
Dispositions for the pr-verify fail note and both pr-review-main verdicts on 510a9cd. Fixes are on df7973a (commits d957fd4, 44a147f, 92d2081, 96ab0a5, 3f91aaa, f83cb1f, 59d9c5d, df7973a, on a merge of main). pr-verify
pr-review-main: claude-review and claude-security-review
Sibling sweep, third pass: every candidate was opened by hand. 24 more sites moved to the enabled wording or the enable-then-install remedy, across architecture, work-items, plugin-quality, overengineering, source-control, implementation, planning, knowledge, review, testing and the loop-lane contract. Left as not gates: tool and agent installs, install-state text, descriptions, reconfiguration copies, plugin-sync text, and the standing "prefer installed skills" instruction. Checks on the new head: |
Closes #16430
Siblings: #16446
Summary
Completes item #16430 on top of #16446, which already landed the seam-phrasing contract change, the hook-unit rule, the handoff "Plugin profile" sentence, the course-digest edit and the handoff eval. This PR adds what #16446 left: the marketplace-wide sweep of gate sites from "installed" to "enabled", the remaining scan edits, the checks that follow the contract wording, and ADR 0109 recording the plugin-profiles one-way doors.
Fix
docs/plugin-philosophy.md's rule line and the.claude/rulessuccessor rule.plugins/session-flow/reference/structure.md: the save-point's suggested-skills slot says "if enabled" (the "Plugin profile" sentence is feat(session-flow): gate optional plugin calls on available skills and record the plugin profile in handoffs #16446's).workflow/context/continuation.mdgates/discovery:read-docs(fallback WebFetch);orientgates/visualization:visualize(fallback an inline mermaid block);planning/plangates/discipline:pick-for-the-problemand/performance:goalwith an inline fallback, and its## Nextdrops "runs in this session", which a planning profile without the implementation plugin makes false.docs/contracts/loop-lane/README.mdcarries the same two wording changes.plugin-quality'scollect-standards.shaccepts an enabled gate and still grades a convention worded "explicit installed-ness";skill-quality's## Nextcheck flagsis|if|when enabledas whole words, each with a test.docs/adr/0109-switch-plugin-sets-by-profile-and-gate-optional-calls-on-enabled-skills.md: profile format and keys, profile homes, extend rule, marketplace governance, the two hook units,coreremoval, the seam gate, the Phase 1 probe record, and the owner's Q75 answer (harness-ops does not depend on session-telemetry).plugins/planning/tests/interview-defenses.test.sh: two section digests re-pinned for the gate wording in the planning interview's SKILL.md Step 4 and loop.md relentless mechanics; no defense sentence changed.Wording basis: the merged contract allows "an equivalent ... clause" and itself uses "when the
other-pluginplugin is enabled" (docs/contracts/seam-phrasing/README.md, "The shape" item 1), so swept sites read "(if that plugin is enabled)" / "(if enabled)".Gate-site classification
Swept every tracked markdown file for installed-ness phrasing (
plugin is installed,(if installed),when installed,when it is installed,when X is (not) installed,skill/capability is installed) and edited gate sites only.**/evals/**,**/fixtures/**, script testsdocs/adr/**,CHANGELOG.mdreference/plugin-reconfiguration.mdcopies and their contractartifact-protocol.mdcopiesharness-ops/skills/plugins/**,audit-skill-visibility,audit-mod-posture, render-report remediesplaygrounds/**,native-playground.mdpresence lines,installed from its marketplacedocs/catalog.md,plugin.jsonand SKILL.md frontmatter descriptionsdocs/upstream/**,vendor/**, skill-authoring's paraphrase of upstreamwhether or not ... is installed,(not installed)state labels,run-e2e's(app-control not installed)report stringVerification
bash plugins/skill-quality/scripts/check-skill.test.sh: all assertions passed after merging main.bash plugins/plugin-quality/skills/audit/scripts/collect-standards.test.sh: all passed.uvx pytest plugins/session-flow/scripts/tests/test_save_point.py plugins/session-flow/scripts/tests/test_workflow_continuation_rule.py: 290 passed.bash plugins/planning/tests/interview-defenses.test.sh: PASS=185 FAIL=0 on the merged head.check-affected: passed on the merged head.plugins/session-flow/evals/ho-disabled-profile-plugin-no-entrypassed 4 of 4 runs (Gate optional plugin calls on enabled skills and record the profile in handoffs #16430 (comment)); this branch's own workflow eval was dropped as a duplicate.docs/review/skills-and-agents.md,plugins/playbooks/skills/skill-authoring/reference/skill-criteria.md"A cut to an existing skill ships only when its evals show no loss").Related
Refs #16428, #16446
Follow-ups
🤖 Generated with Claude Code
https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ