Agent capability-profile migration

The core agent sources declare CLI-agnostic capability profiles (metadata.model:) instead of a Claude Code model alias (metadata.claude.model), per spec 0200 and its delta-01. This is the HOW record of that change: the per-source translation table, the adopter-facing migration note, the corrected tiering audit, the narrowing this change set's own requirement 22 forces on two of the parent spec's requirements, and the observable diagnostic-stream change. The authoring-facing account of what a capability profile is and how a source declares one lives in docs/authoring.md.

The 23-row translation table

<!-- crewrig-table: agent-profile-baseline -->

23 rows: 22 migrated sources plus harness-curator, the tree's one profile-less agent source (requirement 5). "prose, <alias-or-offering>" means the compiled description gains the sentence naming that alias or offering, appended after a single space, with no new frontmatter key (requirements 14 and 18). "unchanged" on GitHub Copilot CLI means the compiled output is identical to its content at 723ad8f except for the metadata.provenance.version line every source's requirement 22 bump reaches (requirement 17 as narrowed by delta-01). harness-curator's four "unchanged" cells mean byte-identical to 723ad8f on every target (requirement 20) — it declares no profile and keeps session-model inheritance (spec 0195 requirement 3).

Adopter migration note

An organization that has forked this framework and carries its own agent sources under artifacts/community/ or artifacts/org/ migrates them off a CLI-namespaced model declaration the same way this change set migrates the 22 core sources: replace metadata.claude.model: <alias> with metadata.model.intelligence: <rung>, using the anchor table of spec 0195 requirement 7 to pick the rung its existing alias corresponds to.

What changes which model a declared rung resolves to is never an edit to an agent source. It is the organization-level override channel of spec 0199 — model-mappings/<target>.org.yml, documented at docs/org-model-mapping-override.md — merged into the mapping in force before a build resolves a source's profile against it. A fork that wants medium to select a different Claude Code offering than the core mapping's default changes model-mappings/claude.org.yml, never metadata.model.intelligence on any agent source.

A fork that declares no profile on its own agent sources and populates no override channel file needs to take no action. Its own agent sources keep exactly the behavior they have today — session-model inheritance — and its own compiled outputs for those sources are unaffected by this change set synchronizing from upstream. This is the same no-action property harness-curator witnesses in the tree above.

Corrected tiering audit (requirement 31)

The change set was audited, case-insensitively, for every document that presents a Claude Code model alias (haiku, sonnet, opus) as the way to choose an agent's model, over the whole tree (excluding .git/ only), tests/e2e/ included. Seven classes of hit exist, and — the point of the audit — none of the seven presents an alias as the way to choose an agent's model, beyond the four surfaces requirements 25 through 28 of spec 0200 already name and correct.

Conclusion — unchanged from the audit's first pass, and now exhaustively grounded. Classes 2 and 6 are mapping and fixture material, class 3 is a test harness's LLM backend, class 4 is a subject under test, class 5 is a CI action input, and class 7 is the normative anchor table and the mappings that declare their own offerings on it — a spec quoting the anchor table and a mapping declaring its offerings are the two places an alias legitimately belongs, and neither presents one as an authoring choice.

This change set's own footprint. Two things a re-audit after this merge will find that were not present before it, both by design, neither contradicting the conclusion above:

• .claude/agents/ goes from 0 to 22 hits. Every migrated agent's compiled Claude description now ends with Run this agent on the <alias> model. — the resolution's own emission under requirement 14, not an authoring instruction. The other three compiled agent trees stay at 0 hits: the Antigravity guidance prose names a Gemini offering, not a Claude alias.
• This change set's own new and edited test files add further, deliberate alias mentions, in the same spirit as class 6 above: the refreshed C2(b) literal of test-model-resolution.sh now pins architect's compiled description verbatim, which itself contains opus; the new test-check-component-metadata-keys.sh declares metadata.claude.model: sonnet in four fixtures to exercise the legacy-key rejection. The migration test suite this change set also adds (scripts/tests/test-agent-profile-migration.sh) asserts the compiled guidance prose of real migrated sources and will add further such mentions for the same reason. None of these is a document presenting an alias as the way to choose a model — every one is a test asserting what the build does, which is exactly class 6's existing category.

Two statements in docs/cli-matrix.md are falsified by this change set and corrected in the same diff as this document: row 33's "nothing in the repository reads it yet" (requirement 27, part one — false since spec 0198 merged), and the [GAP] note asserting that build-components.sh omits model: frontmatter from every compiled agent file (requirement 27, adjacent accuracy — narrowed to the two surfaces Copilot CLI actually reads, .claude/agents/ and .github/agents/, where it stays true). Neither presents an alias as an authoring choice, so neither is a requirement-31 hit; both are named here so the audit is not read as having missed them.

The narrowing of requirements 17 and 19

Spec 0200 delta-01 is the normative source for this narrowing — not a plan reading. Requirement 22 obliges every source this change set modifies to bump its metadata.provenance.version by a MINOR increment, and that bump reaches all four of a migrated source's compiled outputs through inject_provenance (three targets) and gemini_provenance_comment (the fourth), which made the parent spec's original requirements 17 and 19 unsatisfiable in conjunction with requirement 22. Delta-01 replaces both:

• Requirement 17, replaced. Each migrated source's compiled GitHub Copilot CLI output SHALL be identical to its content at 723ad8f except for the metadata.provenance.version line requirement 22 obliges, and SHALL differ in no other byte. Measured on this change set: 22/22 .github/agents/*.md outputs differ from 723ad8f by exactly that line; 0/22 carry any other changed line.
• Requirement 19, replaced. The compiled body of every agent output on every target SHALL be byte-identical to its content at 723ad8f, the provenance carrier excluded from the body for this purpose — the metadata.provenance block on Claude Code, GitHub Copilot CLI and Antigravity CLI, and the <!-- crewrig-provenance: … --> line on Gemini CLI. Measured: 0/22 compiled agent bodies differ on each of the four trees once the carrier is excluded; without the exclusion, 22/22 differ on .gemini/agents/ (the carrier there sits between the closing frontmatter fence and the body) and 0/22 on the other three (the carrier there sits inside the frontmatter, outside "body" under any reading).

Decision 6's third clause is corrected. The parent spec's rationale originally read "the four GitHub Copilot CLI agent outputs stay byte-identical to 723ad8f" — the migration covers 22 sources, each producing one compiled GitHub Copilot CLI output, so the correct count is 22, not four (four is correct for harness-curator's per-target output count, one per target across four targets, which is a different quantity the original clause's four appears to have been carried over from). The corrected clause: the 22 GitHub Copilot CLI agent outputs stay identical to 723ad8f except for the metadata.provenance.version line requirement 22 obliges, and differ in no other byte.

The diagnostic stream

Measured on the migrated tree with all four targets built:

Mechanism. --check ends by running a second build under a fresh mktemp -d root (scripts/tests/test-assembly-verification.sh) that holds no model-mappings/, so the resolution finds no mapping in force and emits a no-mapping note per profiled agent per target. This is a property of --check's own second, synthetic build — not of "the migrated tree" requirement 21 binds — so requirement 21 is satisfied exactly by the 44-line stream a plain build emits, and the 88 additional lines are a real, new-since-this-change-set observable on every CI run of the component-drift capability (which invokes --check): exit status stays 0, and test-assembly-verification.sh still passes, since it verifies assembly structure rather than model emissions. No requirement of spec 0200 binds that suite's synthetic root, and giving it one would change a check that verifies assembly for every tier and every component — a blast radius wider than this change set, left to a follow-up ticket if a reviewer wants it closed.

Decoupling record (requirements 32-34)

No requirement of spec 0200 is conditioned on the outcome of probe C of issue #1113, and no source this change set declares carries a reasoning axis, so every verdict of that probe is absorbable by a change to a mapping file alone (requirement 32). Retiring a target's guidance surface in the future requires no change to any migrated source either: the sources declare needs, and the mapping in force decides the surface a need is stated on (requirement 33).

This change set lands in the nested .claude/agents/<name>/AGENT.md compiled layout. Seam (g) of epic #1100 — the flat .claude/agents/<name>.md layout — is a later, separate ticket; when it lands, it moves the files this change set regenerates to a new path and regenerates the same content there. It re-decides no profile this change set declares (requirement 34).

Follow-up (spec 0201, issue #1127). The move requirement 34 anticipated has happened: seam (g) landed the flat .claude/agents/<name>.md compiled layout, regenerating the same 22 sources' content at the new path rather than re-deciding any profile. An adopter who deployed the retired nested layout into ~/.claude/agents/<name>/ before this move has the stale directory removed automatically, with no manual action, at their next assisted Claude Code setup — the setup's install_tier_to_home now installs each agent as a flat <name>.md file and removes any same-name directory already present. A synchronizing fork lands on the new layout without acting either: the compiled agent trees carry the regenerable policy (.crewrig/core-paths.txt), so the next scripts/sync-from-upstream.sh run restores the new flat files, removes the retired nested ones as orphans, and reports rather than aborts on any member a fork had locally diverged.