refactor(harness): declare tool layout in one adapter table

.agents/shared.md

@@ -45,6 +45,7 @@ self-hosts: this repo's own `openspec/` tree is managed by cospec.
 cospec/
 ├── apps/cli/          @aligned-team/cospec — the cospec CLI
 │   ├── src/canon/     single source of truth: schemas + workflows + gate
+│   ├── src/harness/   HARNESS_TABLE (per-tool layout) + the harness renderer
 │   ├── src/core/      openspec wrapper, parsers, validation, managed files
 │   ├── src/commands/  one file per cospec subcommand
 │   └── test/          unit / contract / integration / fixtures
@@ -249,9 +250,25 @@ fails, fix the root cause; never use `--no-verify`, `pre-commit`, or raw
 
 **Managed files are generated** — `openspec/schemas/**` and the harness dirs
 (`.claude/`, `.agents/skills/cospec-*/`, `.codex/`, `.opencode/`) are composed
-from `apps/cli/src/canon/`. Edit the canon, run `mise run generate`; never
-hand-edit generated output. The `generate:check` drift gate blocks the commit
-otherwise.
+from `apps/cli/src/canon/` (schemas, workflow bodies and workflow identity) and
+`HARNESS_TABLE` in `apps/cli/src/harness/adapters.ts` — the one declaration of
+each tool's layout: skills and commands dirs, filenames, serializer,
+frontmatter, body dialect, rules file, detection paths and receipt note.
+`render.ts`, `init`, `update` and `doctor` all read the table; none keeps its
+own copy of a layout fact. A new tool is mostly a new row, not only one: a
+home-scoped skills root renders but is not yet written; the legacy-skills
+migration (`harness/legacy-skills.ts`, its receipt and `update --check` lines,
+doctor's `legacy-layout` warning) covers only Codex's `.codex/skills`, so a new
+row's `legacySkillsDirs` is detected but never migrated; and deliberate
+Claude-only behaviour sits outside the table — `init`'s `.claude/settings.json`
+merge and its `claude` default (docs/harness-integration.md names them). The
+receipt's `/cospec:propose` hint, always in Claude's spelling regardless of
+selected row, and doctor's scan reading every `.md` file under a skills root
+rather than just `SKILL.md` and the table's command paths, are known defects on
+`main`, not part of that deliberate set; the follow-on change
+`harness-receipt-and-doctor-scope` fixes both. Edit the canon or the table, run
+`mise run generate`; never hand-edit generated output. The `generate:check`
+drift gate blocks the commit otherwise.
 
 **Error handling** — never silently swallow errors. Catch only specific expected
 cases; let unexpected exceptions propagate. Fixes must change observable

.prettierignore

@@ -31,3 +31,11 @@ openspec/schemas/
 # checked-in fixture itself must never be reformatted, or the scenario would
 # start pre-completed.
 packages/bench/scenarios/fixtures/style/src/format.ts
+
+# harness-adapter-table (R8): committed raw golden files proving
+# renderHarnessFiles/init/update/doctor output is byte-identical across the
+# refactor (Buffer-compared, regenerated only under COSPEC_GOLDEN_WRITE=1).
+# `renderHarnessFiles`'s own output is their formatting authority; reflowing
+# them here would make a committed golden diverge from what the code actually
+# emits, defeating the byte-identity proof.
+__golden__/

AGENTS.md

@@ -49,6 +49,7 @@ self-hosts: this repo's own `openspec/` tree is managed by cospec.
 cospec/
 ├── apps/cli/          @aligned-team/cospec — the cospec CLI
 │   ├── src/canon/     single source of truth: schemas + workflows + gate
+│   ├── src/harness/   HARNESS_TABLE (per-tool layout) + the harness renderer
 │   ├── src/core/      openspec wrapper, parsers, validation, managed files
 │   ├── src/commands/  one file per cospec subcommand
 │   └── test/          unit / contract / integration / fixtures
@@ -253,9 +254,25 @@ fails, fix the root cause; never use `--no-verify`, `pre-commit`, or raw
 
 **Managed files are generated** — `openspec/schemas/**` and the harness dirs
 (`.claude/`, `.agents/skills/cospec-*/`, `.codex/`, `.opencode/`) are composed
-from `apps/cli/src/canon/`. Edit the canon, run `mise run generate`; never
-hand-edit generated output. The `generate:check` drift gate blocks the commit
-otherwise.
+from `apps/cli/src/canon/` (schemas, workflow bodies and workflow identity) and
+`HARNESS_TABLE` in `apps/cli/src/harness/adapters.ts` — the one declaration of
+each tool's layout: skills and commands dirs, filenames, serializer,
+frontmatter, body dialect, rules file, detection paths and receipt note.
+`render.ts`, `init`, `update` and `doctor` all read the table; none keeps its
+own copy of a layout fact. A new tool is mostly a new row, not only one: a
+home-scoped skills root renders but is not yet written; the legacy-skills
+migration (`harness/legacy-skills.ts`, its receipt and `update --check` lines,
+doctor's `legacy-layout` warning) covers only Codex's `.codex/skills`, so a new
+row's `legacySkillsDirs` is detected but never migrated; and deliberate
+Claude-only behaviour sits outside the table — `init`'s `.claude/settings.json`
+merge and its `claude` default (docs/harness-integration.md names them). The
+receipt's `/cospec:propose` hint, always in Claude's spelling regardless of
+selected row, and doctor's scan reading every `.md` file under a skills root
+rather than just `SKILL.md` and the table's command paths, are known defects on
+`main`, not part of that deliberate set; the follow-on change
+`harness-receipt-and-doctor-scope` fixes both. Edit the canon or the table, run
+`mise run generate`; never hand-edit generated output. The `generate:check`
+drift gate blocks the commit otherwise.
 
 **Error handling** — never silently swallow errors. Catch only specific expected
 cases; let unexpected exceptions propagate. Fixes must change observable

CLAUDE.md

@@ -45,6 +45,7 @@ self-hosts: this repo's own `openspec/` tree is managed by cospec.
 cospec/
 ├── apps/cli/          @aligned-team/cospec — the cospec CLI
 │   ├── src/canon/     single source of truth: schemas + workflows + gate
+│   ├── src/harness/   HARNESS_TABLE (per-tool layout) + the harness renderer
 │   ├── src/core/      openspec wrapper, parsers, validation, managed files
 │   ├── src/commands/  one file per cospec subcommand
 │   └── test/          unit / contract / integration / fixtures
@@ -249,9 +250,25 @@ fails, fix the root cause; never use `--no-verify`, `pre-commit`, or raw
 
 **Managed files are generated** — `openspec/schemas/**` and the harness dirs
 (`.claude/`, `.agents/skills/cospec-*/`, `.codex/`, `.opencode/`) are composed
-from `apps/cli/src/canon/`. Edit the canon, run `mise run generate`; never
-hand-edit generated output. The `generate:check` drift gate blocks the commit
-otherwise.
+from `apps/cli/src/canon/` (schemas, workflow bodies and workflow identity) and
+`HARNESS_TABLE` in `apps/cli/src/harness/adapters.ts` — the one declaration of
+each tool's layout: skills and commands dirs, filenames, serializer,
+frontmatter, body dialect, rules file, detection paths and receipt note.
+`render.ts`, `init`, `update` and `doctor` all read the table; none keeps its
+own copy of a layout fact. A new tool is mostly a new row, not only one: a
+home-scoped skills root renders but is not yet written; the legacy-skills
+migration (`harness/legacy-skills.ts`, its receipt and `update --check` lines,
+doctor's `legacy-layout` warning) covers only Codex's `.codex/skills`, so a new
+row's `legacySkillsDirs` is detected but never migrated; and deliberate
+Claude-only behaviour sits outside the table — `init`'s `.claude/settings.json`
+merge and its `claude` default (docs/harness-integration.md names them). The
+receipt's `/cospec:propose` hint, always in Claude's spelling regardless of
+selected row, and doctor's scan reading every `.md` file under a skills root
+rather than just `SKILL.md` and the table's command paths, are known defects on
+`main`, not part of that deliberate set; the follow-on change
+`harness-receipt-and-doctor-scope` fixes both. Edit the canon or the table, run
+`mise run generate`; never hand-edit generated output. The `generate:check`
+drift gate blocks the commit otherwise.
 
 **Error handling** — never silently swallow errors. Catch only specific expected
 cases; let unexpected exceptions propagate. Fixes must change observable

apps/cli/src/canon/workflows/harness.yaml

@@ -1,21 +1,10 @@
-# Per-harness rendering manifest (DESIGN 6.1). Consumed by src/harness/render.ts.
+# Workflow identity manifest (DESIGN 6.1). Consumed by src/harness/render.ts.
 # Workflow bodies are single-sourced in the sibling <id>.md files; render.ts pairs each
-# workflow with each selected harness surface and stamps generatedBy/contentHash frontmatter.
+# workflow with each selected tool row and stamps generatedBy/contentHash frontmatter.
 #
-# Skill directories exist for every harness. Command (slash) surfaces exist for claude and
-# opencode only; codex gets skills plus a prefix-rule allowlist (rulesPath) and no slash
-# commands. Path templates use {command} and {skill} placeholders.
-#
-# `codex` and `agents` both render into the vendor-neutral `.agents/skills` root that
-# Codex, Zed, Antigravity and other AGENTS.md-aware assistants read. They share one
-# bodyDialect, so selecting both emits one byte-identical set of files plus codex's
-# rules file. `legacySkillDirs` records the pre-1.11 location cospec migrates away from.
-#
-# `bodyDialect` picks how in-body `/cospec:<id>` references are respelled:
-#   canonical - left as `/cospec:<id>` (Claude registers those slash commands)
-#   opencode  - `/cospec-<id>`, matching the slash commands OpenCode registers
-#   shared    - `$cospec-<skill> (Codex) or /cospec-<skill> (other agents)`, because the
-#               shared root emits no command files and only skill names resolve there
+# This file declares workflow identity only. Every tool's layout — skills and commands
+# roots, filename templates, body dialect, rules file, legacy dirs, detection paths — is
+# declared once, in HARNESS_TABLE in src/harness/adapters.ts.
 #
 # `takesArguments` is audited per workflow: true when the body reads a positional
 # argument (a `<type>: <text>` or a change slug). It drives the OpenCode $ARGUMENTS
@@ -124,23 +113,3 @@ workflows:
     description: >-
       Walk a first-time user through one real cospec change end to end, narrating each step.
       Also use when the user says "cospec onboard" or "openspec onboard".
-
-harnesses:
-  claude:
-    commandDir: .claude/commands/cospec
-    commandFile: '{command}.md'
-    skillDir: .claude/skills/{skill}
-    bodyDialect: canonical
-  codex:
-    skillDir: .agents/skills/{skill}
-    legacySkillDirs: ['.codex/skills/{skill}']
-    rulesPath: .codex/rules/cospec.rules
-    bodyDialect: shared
-  agents:
-    skillDir: .agents/skills/{skill}
-    bodyDialect: shared
-  opencode:
-    commandDir: .opencode/commands
-    commandFile: 'cospec-{command}.md'
-    skillDir: .opencode/skills/{skill}
-    bodyDialect: opencode