@erclx/canon 4.0.0

package/claude/skills/canon-feedback-triage/SKILL.md

@@ -0,0 +1,63 @@
+---
+name: canon-feedback-triage
+description: Triage open GitHub issues labeled `feedback` in the toolkit repo. List them, pick one, classify it as a direct fix or plan-worthy work, route to the right skill, and link the issue for close-out. Use when asked to "triage toolkit feedback", "work through the feedback issues", "process feedback issues", or "what feedback is open". Do NOT use to file new feedback (that is `canon-feedback-file`), or for general GitHub issue triage unrelated to toolkit feedback.
+---
+
+# Canon feedback triage
+
+Consume the feedback queue that `canon feedback --github` fills. Turn an open `feedback` issue into a scoped fix or a plan, then link the issue so merge closes it.
+
+Run from the toolkit repo root. This skill reads GitHub issues, not local `.claude/review/` files. Those are ephemeral session scratch. The durable, cross-project queue is GitHub.
+
+## Guards
+
+- If `gh` is not on PATH, stop: `❌ gh CLI not found. Install it to read feedback issues.`
+- If `gh auth status` fails, stop: `❌ gh is not authenticated. Run gh auth login.`
+- If no open `feedback` issues exist, stop: `✅ No open feedback issues.`
+
+## Step 1: list the queue
+
+Fetch open feedback issues:
+
+```bash
+gh issue list --label feedback --state open --json number,title,url --jq '.[] | "#\(.number)\t\(.title)\t\(.url)"'
+```
+
+Print each as a numbered line with its issue number, title, and URL. Do not open bodies yet. Ask which to triage, and accept "all" to work them in order.
+
+## Step 2: read and classify
+
+For each picked issue, read the body:
+
+```bash
+gh issue view <n> --json title,body,labels --jq '"\(.title)\n\n\(.body)"'
+```
+
+Classify against the toolkit's own surfaces. Score in this order and stop at the first match:
+
+1. **Direct fix.** One surface, one file, no architectural choice. A typo, a stale reference, a one-line rule or doc correction, a single wording fix. Route straight to the edit.
+2. **Plan-worthy.** Multiple files, a new skill or rule, a behavior change, or a cross-surface move. Route to `claude-feature`.
+3. **Needs clarification.** The observed and expected behavior conflict or the surface is unnamed. Comment on the issue asking for the missing detail, then skip it.
+
+State the class and the one-line reason per issue before routing. Do not batch unrelated fixes into one branch.
+
+## Step 3: route
+
+- Direct fix: rename the branch to a conventional name (invoke `git-branch`), make the edit, then open the PR with `git-pr`.
+- Plan-worthy: invoke `claude-feature` with the issue body as the feature description. Let it write the plan and stop. Hand the plan slug back to the user. Do not implement.
+- Needs clarification: `gh issue comment <n> --body "<one question>"`, then move on.
+
+Match one issue to one branch and one PR. A single feedback issue is a single unit of work.
+
+## Step 4: close-out
+
+Link every fix back to its issue so the queue drains on merge.
+
+- For a PR-backed fix, add a `Closes #<n>` line to the PR body so GitHub closes the issue on merge. When `git-pr` regenerates the body, keep that line.
+- For a fix that ships without a PR, close it directly: `gh issue close <n> --comment "Fixed in <commit or PR url>."`
+- For a plan-worthy route, leave the issue open. It closes when the resulting PR merges with its `Closes #<n>` line.
+
+## Notes
+
+- The `feedback` label is what `canon feedback --github` and the `toolkit-feedback.yml` issue form both apply. An issue without it does not surface here by design.
+- This skill routes, it does not reimplement. `claude-feature` owns planning, `git-pr` owns the PR body, `git-branch` owns the branch name. Do not duplicate their logic.

package/claude/skills/canon-operator/REQUIREMENT.md

@@ -0,0 +1,61 @@
+---
+name: canon-operator
+description: Why the toolkit has one front door, what orienting on live catalogs prevents, and where it hands off
+---
+
+# Canon operator requirement
+
+## Gap
+
+Without this skill, using the toolkit means knowing which skill to pick, and the user who most needs it is the one who cannot. A plain-language intent has no obvious owner among a dozen setup and sync skills, so the session guesses a command instead of resolving one, and a wrong guess installs into a project rather than reporting a mismatch.
+
+Two failures come from acting on memory. A session that names a stack, rule, or snippet from what it remembers rather than from the live catalog runs against an entry that has been renamed or removed. And a session that edits a managed file by hand produces a change the next sync overwrites, which reads as the toolkit undoing work rather than as the edit having been made in the wrong place.
+
+The third failure is duplication. A front door that answers everything itself reimplements first-time detection and seed diffing inline, badly, beside the skills that already do both. The value of a router is that it stops at the handoff.
+
+A fourth failure sits between the router and what it runs. This skill executes the install and sync commands whose effect on an existing file is invisible from the command line, and the reference stating that effect per surface is named in the boundary below, which Claude Code never loads. The router therefore overwrites a customized golden config with the answer sitting one skill away and unread.
+
+The fifth failure is writing without re-checking. A domain sync brought a stricter standard into a target and left ten of its eleven context entries non-compliant in the same moment, and the run reported success because reporting the command was where its work ended. A person found the breakage afterwards by running an audit by hand. The audit that would have caught it is built and reachable from nothing the router offers, so one session held both the defect and its detector and connected neither to the other.
+
+The last failure is a section no route reaches. `## Route` maps an intent or a diagnostic finding to one lifecycle phase, so audit offers sitting below that table with no row naming them are unreachable by a correct reading. Two runs asked what the toolkit could measure and both took the scaffold handoff, the row the table declares. A route in settles half of it, since a project carrying a context folder and no installed domain fires the scaffold row and two audit conditions at once, and a body ranking neither leaves the choice to whichever the session read last.
+
+## Must
+
+- Orient on the toolkit's own docs and the live catalogs before acting
+- Map the stated intent to one lifecycle phase, then either run the simplest command that satisfies it or hand off
+- Resolve every stack, rule, snippet, and standard name from a catalog at runtime
+- Run the CLI non-interactively and report the command run, what changed, and the full path of anything written
+- Read the overwrite contract from the body before any install or sync, and name each surface it lists as overwritten before running one
+- Read the reverse walk beside the forward sections, so a folder the toolkit stopped shipping and a migration with a live case both reach the user
+- Hand off a named migration to the skill the report names, and name an unmigrated domain's root and install path for the user to move themselves when no skill covers it
+- Route a measurement intent to the audit offers from the route table itself, rather than leaving the section reachable only by reading past the table
+- Rank a lifecycle row against the audit offers where the route table itself states it, so a session acting on the lifecycle row reads the ranking without opening the section being ranked
+- Offer every audit whose surface the target carries, and offer none whose surface it lacks
+- Hand off an audit that is a skill rather than a command, since the execute contract governs a CLI run and says nothing about a skill's own guards
+- Re-check the target after any operation that wrote, and report what those writes changed against the state read before acting
+
+## Must not
+
+- Edit a managed file by hand instead of running the CLI that owns it
+- Reimplement a flow another skill owns
+- Hardcode a catalog name
+- Move, delete, or name a command against a folder in `unclaimed`. The toolkit stopped claiming the path and only the user knows whether the project owns what sits there
+- Run an audit the user did not pick, which turns a front door into a full sweep
+- Refuse to finish an operation over an audit finding, since every audit reports judgments beside facts and a router that stops on one is a router a target works around
+- Auto-trigger. It is a door the user opens, and a router that fires on its own routes requests nobody made.
+
+## Guards
+
+- An intent matching a deep flow hands off rather than running a shallow version of it, since a partial scaffold is harder to recover from than none
+- An ignore-only fix on an unmeasured tooling report stops, names which cause applies, and names the command answering that cause. Unmeasured has several causes and they do not share a remedy, so one suggested command is wrong for at least one of them. No cause asks the user to supply a stack name, and none reads the zero counts under an unmeasured report as a clean target.
+- A re-check with no earlier report to compare against says the write ran without a baseline, rather than describing the target from scratch
+- A report missing a field this skill reads names the CLI version as the cause and reports that section as unread. Neither an absent key nor a hand-rolled substitute for it may be reported as a finding, since the first reads as a clean target and the second reaches a verdict only the field's attribution can make
+
+## Out of scope
+
+- First-time scaffold of a fresh project: `setup-init`
+- Seed and preamble drift in installed files: `claude-seed-sync`
+- Governance rule install and index bootstrap: `setup-gov` and `setup-indexes`
+- What a given sync overwrites once it runs: `canon-cli`
+
+That last contract carries no row for the governance install this skill routes to, so the overwrite `Must` above reads nothing at one of its own destinations. The row belongs to the skill owning the table. This file records the silence rather than answering it, since a rule written here for a silent contract teaches a session that silence means safe.

package/claude/skills/canon-operator/SKILL.md

@@ -0,0 +1,108 @@
+---
+name: canon-operator
+description: Front door to the toolkit in a target project. Orients on the toolkit's own docs and live `canon` catalogs, then runs or routes any toolkit operation from a plain-language intent. Use when you want one entry point instead of picking a specific setup or sync skill, or when asked to "use the toolkit", "what can the toolkit do", "sync my standards", "install rules", or "help me set up this project". User-invoked only. Defers first-time scaffold to setup-init and seed drift to claude-seed-sync.
+disable-model-invocation: true
+---
+
+# Canon operator
+
+Front door to the toolkit. Orient first, then run the simplest operation that satisfies the intent or hand off to the owning skill. Never bypass `canon` to edit managed files by hand.
+
+## Orient
+
+Run these from the target project root, in parallel, before acting:
+
+- `canon docs list`: the index of toolkit reference docs
+- `canon docs agents`: the CLI command catalog and invocation contract
+- `canon docs target-projects`: the scaffold, add-a-domain, and sync lifecycle
+
+Load a domain doc with `canon docs <topic>` only when the intent touches that domain. Read the live catalog for any domain you act on with `canon <domain> list --json`. Never hardcode stack, rule, snippet, or standards names.
+
+## Diagnose
+
+Run `canon sync --check . --json` before routing. It reports what a target is behind on across every surface, so the intent comes from the project state rather than from the user having to know it already. Skip only when the user named a single operation to run.
+
+Read seven fields off the report and carry each to `## Route`:
+
+- `unmigrated`: a domain sitting at the root layout with nothing under `.claude/`. The most urgent finding, because that domain reports no drift of its own while being entirely behind
+- `superseded`: a file a newer seed folder replaced. No command moves it, since the content is the project's own, so this routes to the skill that proposes the split rather than ending at the report
+- `seeds`: entries are `matching`, `stale`, `drifted`, or `missing`. Anything but `matching` needs the seed handoff
+- `domains[].entries`: per-file `stale`, `customized`, `stranded`, and `orphaned` as before
+- `reverse`: what the target holds that no live catalog claims. `unclaimed` lists folders at roots the toolkit stopped shipping, each with an `attribution` of `dropped`, `unattributed`, or `project`. Act on the first two and leave `project` alone, which history proved the project owns. `migrations` names a proposal-only skill with a live case here and the `reason` it was measured from
+- `historyUnavailable` on a domain, on `seeds`, or on `reverse`: nothing could be dated, so treat every difference as unverified and say so rather than reporting a file as untouched. It is set on `reverse` when the toolkit itself ships without history, which is the registry-install case, so the walk found nothing rather than finding a clean target
+- `tooling`: read `measured` first. Every count under it is zero when it is false, which is an absence of measurement rather than a measured zero. Past that, `chain` names the stacks the install resolved, nearest first, and `counts.gitignore` counts the managed ignore entries the target is missing.
+
+State what the report found in one line per finding before acting on any of it.
+
+A `reverse` key the report does not carry at all is a CLI predating the field rather than a target with nothing unclaimed. Say the walk did not run and name the CLI version as the cause. Reading the absent key as an empty answer reports a clean target to a project nobody has looked at, and answering it from a filesystem walk of your own is worse, since that reaches a verdict the attribution the field carries is the only thing entitled to make.
+
+## Route
+
+Map the stated intent, or what `## Diagnose` found, to one lifecycle phase, then act.
+
+The two can name different rows, and a reply answers both rather than picking one. Run or hand off what the finding names, then answer the intent in the same reply. A measurement intent therefore reaches the audit offers below even on a target whose diagnostic found scaffold work, which is the case a session leaving from the scaffold row would otherwise never read.
+
+- A domain in `unmigrated`: name it, the root path, and the install path. No command moves it, so the user runs the relocation themselves
+- An entry in `migrations`: hand off to the skill its `skill` field names, spelled as the report spells it. Both of them propose without writing, so the handoff is where this stops
+- Anything in `superseded`: name which files and what replaced them, then hand off to `migration-superseded`. Both skills propose, so neither moves nor deletes the file
+- A folder in `unclaimed`: name it and the attribution it carries. No command moves it and the content may be the project's own, so the decision is the user's
+
+- First-time scaffold of a fresh project: hand off to `setup-init`
+- Governance rules for the project stack: hand off to `setup-gov`
+- Bootstrap the `index.md` system: hand off to `setup-indexes`
+- Seed or standards drift in `CLAUDE.md` or `.claude/` preambles: hand off to `claude-seed-sync`
+- Install one snippet, standard, or rule: run the domain `install` command
+- Sync one domain or every installed domain: run `canon <domain> sync` or `canon sync`
+- Fix only the ignore entries of the installed stack: run `canon tooling inject --gitignore <stack>`
+- Measure a surface without changing it: offer the audits under `### Audits`
+- Browse what is available: run `canon <domain> list`
+
+### The ignore-only row
+
+The ignore-only row runs on a measured tooling report alone, so `measured` decides before `chain` is read at all. A false one splits three ways. Report tooling as unmeasured in each, name the cause, and run nothing, since the zero counts underneath are unmeasured rather than clean:
+
+- An empty `chain` at a workspace root, which carries `pnpm-workspace.yaml` or a `workspaces` key in `package.json`: no chain is recorded there by design, since one would guess at what the packages hold. Name `canon tooling sync <stack> <path>` against a package.
+- An empty `chain` anywhere else: no tooling install is recorded, so name `canon tooling sync` as the command that records one rather than asking the user for a stack.
+- A `chain` carrying names: this toolkit no longer ships those stacks. Name them, since injecting would write against a retired name.
+
+On a measured report, take `<stack>` from the first name in `tooling.chain`, which records the stack nearest the target. Inject re-resolves that leaf's own chain, so a target whose recorded chain is shorter receives the entries from the layer its install skipped. Say so before running it.
+
+### Audits
+
+Five audits measure a surface without changing it. Offer the ones whose surface the target carries, list them together, and let the user pick. Run none of them unasked, and never treat a finding as a reason to abandon the operation the user asked for, since each reports judgments beside facts.
+
+A lifecycle row and these offers fire together on a project carrying a context folder and no installed domain, which is the ranking the preamble states. Scaffold work on a target with nothing installed is real work, so an audit offered instead of it answers a question nobody asked, while an audit dropped in favor of it loses the one the user did.
+
+- `.claude/context/` present: offer `canon context audit`
+- A record folder present under `.claude/`, one of `plans`, `groundwork`, `intake`, or `memory`: offer `canon records validate <kind>` for each one found
+- Markdown that git lists: offer `canon markdown audit`
+- TypeScript or shell source present: offer `canon comments scan`
+- A `package.json` script that serves an interface, one of `dev`, `preview`, `serve`, or `start`: offer `claude-ux-measure`
+
+An audit offered against a surface the target lacks reports an empty run as a finding, which is the same defect as never offering it at all. Check the surface before naming the command.
+
+The markdown row is the one every target satisfies, since a project with no markdown is not one this reaches. Its condition is stated so the five rows read alike, and the row needs no gate beyond it.
+
+The last row is a skill handoff rather than a command, so `## Execute` does not govern it. Hand it off and stop, the way a lifecycle row hands off. Its condition reads a serving command rather than a source folder, since that skill measures what a browser receives and never opens the tree behind it. It carries a second condition this skill cannot read, which is a browser harness the project already installed, and the skill itself reports what it needs when it finds none. Testing for that here would put the detection in two places and let this one answer stale.
+
+## Execute
+
+For operations this skill runs directly:
+
+- Read the catalog first with `canon <domain> list --json`, then match against project context
+- Read the `canon-cli` skill before any install or sync, and warn the user by name about each surface it lists as overwritten. A command name says nothing about what it does to a file already there, and this skill is the one running it.
+- Run the CLI with `CANON_NON_INTERACTIVE=1` so it skips prompts. The tool permission dialog is the confirmation gate.
+- Report the command run and what changed. Emit the full relative path for any file written.
+- Re-run `canon sync --check . --json` after any operation that wrote, and compare it against the report `## Diagnose` read before acting
+- Report the fields that moved and name the write that moved each one. Do not restate the second report, since the value sits in the difference alone.
+- State a field that moved the wrong way and leave it for the user to decide on. The re-check reports and never repairs.
+
+A write can leave a target worse in a field the write never named, which stays silent while the operation itself succeeds. The re-check is what makes that visible, so it runs before the operation is reported as done.
+
+The comparison needs the earlier report. When `## Diagnose` was skipped because the user named a single operation, say the write ran with no baseline rather than describing the target from scratch, which would attribute differences another session made to this run.
+
+## Boundaries
+
+- Run `canon`. Never reimplement its install or sync logic, and never edit managed files like rules, configs, or seeds by hand.
+- Hand off the deep flows. Do not duplicate `setup-init` detection or `claude-seed-sync` part-diffing inline.
+- Resolve names from catalogs at runtime. A hardcoded name is a bug.

package/claude/skills/canon-rollout/REQUIREMENT.md

@@ -0,0 +1,59 @@
+---
+name: canon-rollout
+description: Why one skill carries both roles of an outbound wave, and where its boundary sits against the seven skills that operate inside a target
+---
+
+# Canon rollout requirement
+
+## Gap
+
+Without this skill, a toolkit change reaches its consuming projects by hand. One measured pass across four targets took four hand-written `cd <path> && claude --bg` invocations, each carrying a brief composed by hand from that project's review. Nothing enumerated the targets, nothing derived the brief, and nothing applied a convention across the set.
+
+The conventions were already settled and reached none of the four. The branch and title shape was answered and reconfirmed, and the four pull requests carried four titles across two scopes, neither of them the settled one, with three naming the toolkit in the half of the shape that withholds it. A convention nothing applies holds only where somebody remembers it.
+
+The return leg dead-ends outside this repository. A review that posts a finding on a target's pull request has to reach the session holding that branch, and every dispatch guessed instead. All four sessions that opened those pull requests were still alive holding their worktrees hours later, two fresh addressers were sent over them, one refused on the worktree lock, and one cut a second worktree on the same branch, which would have put two sessions pushing to one ref.
+
+A wave ends without telling the operator what to do. Reading where four targets stood took a hand-written shell loop over four repositories, re-run three times as reviews landed, and every status line it produced named a repository and a number and no link. Merging is the one act only the operator performs, and they reached each result by navigating to it by hand.
+
+A worker's reply reads as done when it is not. One pull request's fix for two minor findings closed both and introduced four more, all in prose the fix added, including a claim about a workflow that runs on a trigger it does not declare. A loop that closes a target on the worker's word merges that.
+
+## Must
+
+- Enumerate the targets from the record the install writes, with the sweep as the fallback and its stated bound read before any count is trusted, since a hand census of this population was wrong in both directions at once
+- Write the branch and the title from a fixed shape rather than leaving either to the dispatched session, since the settled convention reached none of four hand-driven repairs
+- Carry one line in each target's body naming the toolkit as the source and where to change the content upstream, since a reader in that repository has no other route to the fact that a local edit is lost at the next wave
+- Name no toolkit version in any target's body, since the install stamp in that target already records `syncedAt` and a sha256 per synced file and a version string is the copy that goes stale
+- Dispatch one worker per target, with the branch checked as unclaimed in that target before the launch
+- Branch from the clone that is current rather than from whichever the record names first, since a working copy answers what that machine last pulled rather than what the target holds
+- Resolve who holds a branch before dispatching an addresser, and dispatch a fresh one only where the resolver reports the owning session gone
+- Address every finding whatever its severity, and close a target only when a narrow re-review posts `## Review closed`
+- Name a pull request URL on every line that reports a target, at dispatch and at hand-back both
+- Close a wave by naming every target with its URL, review heading, CI state, and what is owed, the targets needing nothing included
+
+## Must not
+
+- Merge anything, in either role, at any size. The operator holds that gate over every role rather than as a setting on a wave.
+- Reimplement what `canon targets list`, `canon targets pulls`, or `canon sessions list --repository` already answers. Each is a shipped verb this body reads under `CANON_NON_INTERACTIVE=1`.
+- Restate the in-target diagnosis and routing that `canon-operator` owns, or the review, address, and worktree procedures the three skills named below own
+- Read or write this repository's task board. The board is `claude-orchestrate`'s subject and a wave is not a row on it.
+- Bound the review and address loop with a count. The bound was declined with its failure mode stated, and what guards the loop is the orchestrator reviewing every round itself.
+- Edit a managed file in a target by hand rather than through the `canon` verb that owns it
+- Be a skill nothing invokes but its author typing the name. Nothing else routes to it, since it carries `disable-model-invocation: true` and no sibling body names it, so a wave that only ever runs when somebody types the name is the signal that the loop never replaced the hand-driven pass it was built against.
+
+## Guards
+
+- `gh` absent: stop, since no target's pull request can be read or opened
+- The target index reads as unknown with no sweep behind it: stop and name the reason, rather than reporting a population of zero
+- The claim check refuses or comes back with either readability flag false: treat the target as unverified rather than clear and hand it to the operator
+- Every clone of a target is behind its remote: refuse that target rather than branching from a stale base
+- More than one session holds a target's branch: report the ambiguity and stop, rather than sending findings to whichever row came first
+
+## Out of scope
+
+- This repository's own board, queue, and worker dispatch, which `claude-orchestrate` holds
+- The role a session building one branch under one plan in this repository takes, which `claude-worker` holds. A rollout worker does not take it, since that body resolves session scratch against a main worktree root a target does not carry.
+- Entering the worktree, which `claude-worktree` owns, including the branch collision tests it runs before entry
+- Diagnosing what a target is behind on and routing each finding to its command, which `canon-operator` owns
+- Posting a review to a pull request and moving its heading, which `claude-pr-review` owns
+- Answering a posted review inside the target, which `claude-address-review` owns
+- The commit and the pull request mechanics, which the `git-*` family owns

package/claude/skills/canon-rollout/SKILL.md

@@ -0,0 +1,147 @@
+---
+name: canon-rollout
+description: Takes one toolkit change out to every consuming project and brings each to a mergeable pull request. Carries an orchestrator role that enumerates the targets, dispatches a worker into each, reviews every pull request, and routes what each review posts, and a worker role that builds inside one target and answers its review. Use when asked to "roll this out to the targets", "take this change to every project", "run a rollout wave", "update the consuming projects", or when a session was dispatched into a target by a wave. Do NOT use for this repository's own board, which is `claude-orchestrate`, and merge nothing.
+disable-model-invocation: true
+---
+
+# Canon rollout
+
+This skill runs outward. Every other skill for operating on a target assumes the session already sits inside it, and this one starts in the toolkit and reaches the projects the toolkit installed into.
+
+One wave takes one toolkit change to every target and ends with a pull request per target for a person to merge. Two roles carry three phases over one target list. The orchestrator enumerates, dispatches, reviews, and routes. The worker holds one target from the worktree entry to the merge of the branch it opened.
+
+`claude-orchestrate` owns this repository's own board and is a different subject. Read nothing from `.claude/tasks/` here and write nothing to it.
+
+## Take a role before anything else
+
+- The invocation or the launch prompt names the worker role, a target path, or both: worker. Skip to `## The worker role` and run no phase above it.
+- Neither is named: orchestrator. Run the three phases in order.
+
+The role is read off the prompt because nothing else carries it. A dispatched worker starts in the target's own checkout, which is a repository like any other from the session's side, so a test on the working directory answers the same for a worker in a target and an operator who invoked this from one. Reading the role wrong in that direction starts a second wave from inside a consuming project, which is why the dispatch below names the role and the path both rather than relying on either alone.
+
+Do not invoke `canon:claude-worker` from either role. That body states the role for a session building one branch under one plan in this repository, and it resolves session scratch against a main worktree root a target does not carry, so a rollout worker reading it hunts for a plan nobody wrote.
+
+## Guards
+
+- `gh` absent from the path: stop. `❌ gh is not installed, so no target's pull request can be read or opened.`
+- `canon targets list` refuses, or reports the index as unknown with no `--sweep` behind it: stop and name the reason. A count of zero taken from an index nothing read is a confident wrong answer.
+- Read a record's own `reason` field rather than a command's exit status. An operator's shell profile may wrap `canon` in a function whose status comes from a trailing command, which reports every refusal as success.
+- Neither role merges anything, at any size and with no proportionality exception. A wave ends by handing the operator a list to merge from, and merging is the one act only they perform.
+
+## The fixed shape every target receives
+
+Every target in every wave takes the same branch and the same title, written from here rather than chosen per session. Four hand-driven repairs produced four titles across two scopes, neither of them this one, and three named the toolkit in the half of the shape that withholds it.
+
+- Branch: `chore/agents`, in every target and every wave.
+- Title: `chore(agents): <what changed in that target>`.
+- Body: the sections `${CLAUDE_SKILL_DIR}/../../standards/pr.md` fixes, describing what that target received.
+- Body, one line: name this toolkit as the source of the changed files and where to change the content upstream. A reader in that repository months later has no other route to the fact that these files are toolkit-managed and that a local edit is lost at the next wave.
+- Body, no version number anywhere in it. `${CLAUDE_SKILL_DIR}/../../standards/versioning.md` permits a semver tag only where the pull request cuts a release, and `.claude/canon/config.json` in the target already records `syncedAt` and a sha256 per synced file. A version in the body is a second copy of a fact the target holds canonically, and it is the copy that goes stale while the pull request sits open.
+
+The title withholds the source and the body supplies it, because the two have different audiences. `agents` says what changed without saying where it came from, which is what a target's own history wants, and the source line is for a reader inside that repository rather than for whoever merges.
+
+One branch name across every target is also what lets the return leg resolve. `canon sessions list --branch chore/agents --repository <clone>` answers about that project rather than the working one, and a branch each session named for itself gives that read nothing to match.
+
+## Phase 1: enumerate and dispatch
+
+### Read the population
+
+1. Pin the wave. Record the commit of this repository that the wave carries and name it in this session's own reports. Name it in no target's body.
+2. Enumerate. Run `canon targets list --json` under `CANON_NON_INTERACTIVE=1`, adding `--sweep <root...>` to reach targets installed before the index existed. Read the `bound` object back before treating any count as the population, since it names the roots walked, the depth, where the walk stopped, and what it could not read. The bound a sweep can never state is the machine, so a target on another one sits outside every answer it gives.
+3. Pick the clone that is current. A row carries every checkout of that project under `paths` and leads with the one a sync ran in, which records where an install happened rather than what the checkout holds today. Fetch each candidate, compare it against the origin's default branch, and dispatch into one that is level. Refuse one that is behind rather than branching from a stale base, since three clones read as still owing the repair in one census and every one of them was behind its remote and nothing else.
+
+### Check, then dispatch
+
+4. Check the branch is unclaimed in that target. Run `canon sessions list --branch chore/agents --repository <clone> --json` and read `claimed` off the record.
+   - `claimed: true`: something already holds it. Report what does, `worktree` when it names a path, `sessions` when it carries a row, and `refs` when the branch already exists, then leave that target out of the wave rather than colliding.
+   - `claimed: false` with `sessionsReadable` and `refsReadable` both true: dispatch.
+   - Anything else, a refusal or a record carrying no `claimed` key included: treat the target as unverified rather than clear, name which reading could not be taken, and hand it to the operator to launch by hand. Dispatching on a check that could not be read reproduces the collision the check exists to prevent.
+5. Dispatch one worker per target, one session each.
+
+```bash
+cd <clone> && claude --bg --model <model> -n "rollout-<target>" "Run /canon:canon-rollout in the worker role against the target at <clone>. <what the wave carries>"
+```
+
+`--bg` returns immediately, `-n` sets the name that separates a wave's worker from an operator's own launch in the roster, and `--model` overrides the inheritance that otherwise spends this session's model on every worker. Keep the `rollout-` prefix.
+
+Running the targets in turn is the alternative and it puts every target's diff through one session's context. A wave's spend scales with the target count instead, and nothing caps it, so weigh the count before widening rather than after.
+
+Report each dispatch by naming the target, the clone, the branch, the model, and the session name. Carry the pull request URL into every later report from the moment its worker announces one.
+
+## Phase 2: review every pull request
+
+1. Poll. Run `canon targets pulls --json`, naming the wave's clones as arguments to read those alone. A target comes back as `read` with a `pulls` array or as `refused` with a `reason`, and a refusal is not a target with no work. Reading a failed query as no open work reports a target as done having read nothing.
+2. Review here, in this session. Invoke `canon:claude-pr-review` and redirect every call that body makes, since it resolves a pull request from the repository the session stands in and this session stands in the toolkit. Every call takes a redirect, and the ones that silently answer about the wrong repository are worse than the ones that fail:
+   - `--repo <owner>/<name>` on `gh pr view`, `gh pr diff`, and `gh pr review`, where it is a global flag.
+   - `-C <clone>` on `git fetch origin pull/<n>/head`, `git merge-base`, `git diff`, `git log`, `git show`, and `git grep`.
+   - `GH_REPO=<owner>/<name>` on any `gh api repos/{owner}/{repo}/...` call, which is the close-out route. `gh api` takes no `--repo` flag and fills those placeholders from the working directory or from that variable, so the one call a flag cannot redirect is the one that edits a review in place.
+   - That project's own root file, rules, and context, read from under the clone.
+
+   A missed redirect reports rather than fails. A pass that redirects the `gh` half and leaves a `git` call alone reads this repository's history against a target's diff and reports a delta, which is a finding nobody can trace back to the call that produced it.
+
+   This session wrote none of these diffs, so reviewing them costs nothing in independence.
+
+3. Dispatch a reviewer into the target when a diff is too large for this session to hold. That is the fallback rather than the default. Dispatching one per target was measured and what it saved was reading four diffs, not reading four sets of findings, and that second half reaches this session either way, since routing a finding back means composing the brief from it.
+4. Read the heading each pass posted. `canon targets pulls` reports `review` as `open`, `closed`, or `null` when no pass has landed, and reports `reviewReadable` as `false` when the query failed, which leaves `review` covering nothing.
+
+## Phase 3: route every finding and close each target
+
+1. Address every finding whatever its severity. A minor is a finding, and a reviewer calling one non-blocking does not close the pass that raised it.
+2. Resolve who holds the branch before dispatching anybody. Run `canon sessions list --branch chore/agents --repository <clone> --json` and read the count rather than the first row.
+   - A live session holds it: send that session the findings and name `canon:claude-address-review` for it to run, which is step 6 of `## The worker role` and the step that session already stands at. Never name this skill in that message. This body carries `disable-model-invocation: true`, which blocks the `Skill` tool rather than only suppressing an auto-trigger, and a session acting on an inbound message reaches a skill that way and no other, so a message naming this one halts the address leg with the branch built and the findings unread. A launch prompt is the one route to a flagged body, which is why the dispatch above names this skill and this message must not. Assuming the worker was gone is what one hand-driven pass got wrong. All four sessions that opened those pull requests were still alive holding their worktrees hours later, and of two fresh addressers sent over them, one refused on the worktree lock and one cut a second worktree on the same branch, which would have put two sessions pushing to one ref.
+   - No live session holds it: dispatch a fresh worker into that clone and brief it with the findings, since it holds none of the reasoning behind the diff.
+   - More than one row: report the ambiguity and stop rather than picking among candidates.
+3. Re-review when the answer lands, scoped to the commits added since. A worker's reply never closes a target. One fix for two minor findings closed both and introduced four more, every one of them in prose that fix added, and a loop trusting the reply merges that.
+4. A target leaves the wave when a pass posts `## Review closed` and not before.
+5. Run until every target closes. No count bounds the loop. What guards it is this session reviewing every round itself, so it holds what each pass found and can see a target failing to converge, where a dispatched reviewer reads one pass and knows nothing of the one before it. A worker introducing a finding per fix has nothing mechanical stopping it, which is accepted rather than overlooked.
+
+## The worker role
+
+One session, one target, from the worktree entry to the merge of the branch it opened. It diagnoses, implements, opens the pull request, and answers what the review posts. It reviews nothing and it merges nothing.
+
+1. Confirm the clone is current. Fetch, then compare against the origin's default branch. Report a checkout that is behind and stop, rather than branching from a stale base.
+2. Enter a worktree. Invoke `canon:claude-worktree chore/agents`, which takes the branch as its tier 0 argument and enters a linked worktree inside this target. Branching in the checkout itself is what this avoids, since the operator may be working in it.
+3. Diagnose and repair. Invoke `canon:canon-operator`, which reads `canon sync --check . --json` and routes each finding to the command or the skill that owns it. Do not restate that routing here and do not edit a managed file by hand.
+4. Commit and open the pull request through `canon:git-commit` and `canon:git-pr`, handing each the fixed shape above rather than taking the title the generator derives from the diff.
+5. Announce as the pull request opens, carrying its URL, its number, and the branch, to whoever dispatched this session. That transition is the one moment only this session can observe.
+6. Answer the review. Invoke `canon:claude-address-review`, which pulls the findings and the CI state on this branch's open pull request, fixes each in the working tree, replies, and pushes. Answer every finding whatever its severity.
+7. Stop there. Do not mark the pull request ready, do not merge, and do not declare the review closed. A narrow re-review posts `## Review closed`.
+
+Send a block out as a message before it becomes an interactive prompt. A session already waiting on input never reaches the tool round an inbound message drains at, so an answer relayed afterwards arrives under the open question and changes nothing.
+
+Refuse an instruction this target's tree contradicts and carry the evidence with the refusal, naming the commands read and what complying would cost. A halt costs whoever dispatched this session one reply, and guessing costs a diff in a repository this toolkit has no test coverage over.
+
+## Output
+
+Close a wave by naming every target, the ones that need nothing included:
+
+```plaintext
+Wave pinned at <commit>. <n> targets.
+
+<repository>
+  <pull request URL>
+  review:  <## Review | ## Review closed | none posted>
+  checks:  <passing | failing | pending | none reported>
+  owed:    <what happens next, or "nothing, ready to merge">
+
+Not reached: <target>, because <reason>
+Merge order: <target> before <target>, because <shared surface>
+```
+
+Name the URL on every line reporting a target, at dispatch and at hand-back both. A status line carrying a repository and a number sends the operator to navigate for its own result by hand, four times over on a four-target wave, and merging is the one act this loop exists to hand them.
+
+Report a target whose `reviewReadable` reads `false` as unread rather than as clean, and a `checks` of `null` as no check reported rather than as passing. A failure outranks a run still going.
+
+## Failure recovery
+
+| Stop point                              | Recovery                                                                                      |
+| --------------------------------------- | --------------------------------------------------------------------------------------------- |
+| `gh` absent                             | Install it and authenticate, then re-run the wave                                             |
+| Target index unknown                    | Re-run with `--sweep <root...>` and read the `bound` before trusting the count                |
+| Every clone of a target is behind       | Pull that clone, or name a current one, then dispatch that target alone                       |
+| Claim check refuses or reads unverified | Launch that target's worker by hand and report that it went out outside the check             |
+| Branch already claimed in a target      | Resolve what holds `chore/agents` there, then dispatch that target alone                      |
+| A target refuses the pull read          | Read the `reason`, repair it, and re-poll that target. Never record it as having no work      |
+| Diff too large to review here           | Dispatch a reviewer into that clone and route its findings back through Phase 3               |
+| Two sessions hold one branch            | Report both and stop. Sending findings to either puts two sessions on one ref                 |
+| A target fails to converge              | Stop that target's loop, report the rounds it took and what each pass found, and hand it over |

package/claude/skills/canon-screencast/REQUIREMENT.md

@@ -0,0 +1,39 @@
+---
+name: canon-screencast
+description: Why a recording script ships pre-seeded rather than blank, what the four discovery questions buy, and why the draft stays stack-agnostic
+---
+
+# Canon screencast requirement
+
+## Gap
+
+Without this skill, a recording plan arrives as a wall of blank fields, so the user fills a template instead of editing a draft and the work the skill was meant to do lands back on them. The opposite failure costs more. A session that skips discovery guesses the audience, the length, and the hero moment, then writes five beats for a video nobody asked for, and every one of them has to be thrown away rather than edited.
+
+A draft that names the recording software, the editing software, or the window manager pins the script to one machine and one person's setup, so it stops being reusable the moment anyone else records. And re-invoking to refine silently overwrites the edits the user made in the file, which is the one failure that destroys work rather than wasting it.
+
+## Must
+
+- Read project context and recent commits before proposing, since what shipped recently is usually the subject
+- Ask exactly four discovery questions, each carrying a default derived from that context
+- Wait for answers rather than reading silence as acceptance
+- Pre-seed every section with concrete content so the draft is shippable as written and the user edits down
+- Derive the slug from the topic and the discovery answers
+- Write to the gitignored scratch path at the main worktree root and print that path on its own line
+
+## Must not
+
+- Name recording software, editing software, fonts, or window managers
+- Record, edit video, or generate captions
+- Re-run against an existing draft to refine it, since re-running overwrites
+- Infer the hero moment without asking, which is the beat the whole script is built around
+
+## Guards
+
+- No topic provided: stop rather than guessing what is being recorded
+
+## Out of scope
+
+- Producing the recording, which stops at the script by design
+- Refining an existing draft, which is a direct edit of the file
+- Slide decks, which `canon-slides-draft` owns
+- Where the recording ships, which the draft lists and the user decides