refactor(build): compile the Git agent from an MDS generator host

.devflow/features/compliance-feature/KNOWLEDGE.md

@@ -9,7 +9,7 @@ directories:
   - src/cli/commands/compliance.ts
   - src/assets/skills/compliance
   - src/assets/rules/compliance.md
-  - src/assets/agents/git.md
+  - src/assets/agents/git.mds
   - src/assets/commands/code-review.mds
   - src/assets/commands/plan.mds
   - src/assets/commands/implement.mds
@@ -303,7 +303,7 @@ Collects the commit list (≤100 entries) and shipped issue numbers (≤50) sinc
 | `src/core/plugins.ts` | `FEATURE_OWNED_SKILLS`, `FEATURE_OWNED_RULES`, `DELETED_PLUGIN_NAMES`, `resolveFeatureRedirect` |
 | `src/cli/commands/rules.ts` | `seedRuleShadow` (Tier 1 skipped for FEATURE_OWNED_RULES; Tier 2 = canonical source preserves placeholder) |
 | `src/assets/commands/_partials/_compliance.mds` | `compliance_gate()` partial — single-source COMPLIANCE_SKILL_INSTALLED resolution for all 4 host commands |
-| `src/assets/agents/git.md` | All traceability operations (D1–D9 legend, D4 rate-limit backpressure, D9 gate table, gather-release-evidence, setup-task containment, Principle 8 marker neutralisation) |
+| `src/assets/agents/git.mds` (compiles to `dist/agents/git.md`) | All traceability operations (D1–D9 legend, D4 rate-limit backpressure, D9 gate table, gather-release-evidence, setup-task containment, Principle 8 marker neutralisation) |
 | `src/assets/commands/code-review.mds` | Step 0b (imports compliance_gate), Phase 1 regulated-surface gate, Git COMPLIANCE field |
 | `src/assets/commands/resolve.mds` | Phase 1b (fetch-review-threads), Phase 9b (resolve-review-threads), Phase 9c (check-merge-readiness) |
 | `src/assets/commands/plan.mds` | compliance_gate gate for compliance Design agent and mandatory issue linking |

.devflow/features/dynamic-workflow-engine/KNOWLEDGE.md

@@ -18,7 +18,7 @@ directories:
   - dist/commands
   - tests/build-mds.test.ts
 created: 2026-07-07
-updated: 2026-08-22
+updated: 2026-09-09
 ---
 
 # Dynamic Workflow Engine
@@ -68,7 +68,19 @@ Partials declare **no** `output-dir:` frontmatter key. Host files declare it as
 
 ### Compiled output and test pinning
 
-`scripts/build-mds.ts` compiles all 13 host files (9 knowledge + 4 dynamic) — `ALL_HOSTS = 13`. **`DIST_FILES` = 14**: the 13 compiled outputs plus `release.md`, which is hand-authored and copied verbatim by the build; the divergence is permanent (SG-13). Compilation-scope guards use `ALL_HOSTS`; deployed-behaviour guards (gh-issue scope, compliance_gate, retired wording) use `DIST_FILES`. The test file `tests/build-mds.test.ts` reads the compiled `dist/commands/dynamic-build.md` and greps for exact doctrine strings. Changing a doctrine literal in a partial immediately breaks the relevant test — by design. The test suite pins:
+**This KB owns the 13/14/14 count rule.** The `feature-knowledge-system` and `test-harness` KBs point here rather than restating it, so there is one place to correct when a number moves (applies PF-053).
+
+Three numbers, three sets:
+
+| Number | Name | The set it counts | Why it differs from the others |
+|--------|------|-------------------|-------------------------------|
+| **13** | `COMMAND_HOSTS` (`MDS_COMMAND_HOSTS`) | Command hosts under `src/assets/commands/` — 9 knowledge + 4 dynamic — compiled into `dist/commands/` | Excludes `git.mds`, which is a **generator host**: it declares `output-dir: dist/agents` and compiles to `dist/agents/git.md`, never to `dist/commands/` |
+| **14** | `ALL_MDS_HOSTS` | Every host the build discovers and compiles: the 13 command hosts **plus** the `git.mds` generator host | Counts compilation inputs across both destinations |
+| **14** | `DIST_FILES` (`DIST_COMMAND_FILES`) | Files that must exist in `dist/commands/` after a build: the 13 compiled command outputs **plus** `release.md` | `release.md` is hand-authored and copied **verbatim** — it is not a host and is never MDS-compiled; the divergence is permanent (SG-13) |
+
+The two 14s are different sets that happen to share a length: one is inputs (both destinations), one is outputs (one destination). Never conflate the three numbers. Compilation-scope guards use `COMMAND_HOSTS`; deployed-behaviour guards (gh-issue scope, compliance_gate, retired wording) use `DIST_FILES`.
+
+Host and partial names are shared across the suite by the manifest at `tests/fixtures/mds-manifest.ts` (`MDS_COMMAND_HOSTS`, `MDS_GENERATOR_HOSTS`, `MDS_PARTIALS`, `ALL_MDS_HOSTS`, `DIST_COMMAND_FILES`, `HAND_AUTHORED_COMMAND_FILES`) rather than by count literals — the manifest answers "which?", so a rename plus an addition in one commit cannot stay green. `COMMAND_HOSTS` is the local alias `tests/build-mds.test.ts` gives `MDS_COMMAND_HOSTS`. The test file `tests/build-mds.test.ts` reads the compiled `dist/commands/dynamic-build.md` and greps for exact doctrine strings. Changing a doctrine literal in a partial immediately breaks the relevant test — by design. The test suite pins:
 - `Simplify` and `Scrutinize` each appearing exactly **2 times** (Gate 1 #1 + Gate 1 #2 only)
 - **C1 (single-pass review):** presence: `The review pass runs exactly ONCE`, `The pass runs exactly ONCE`, `Never author additional cycles or a delta re-review of fix commits` (invariant #7 unique), `Budget scales roster and verification votes, NEVER the number of passes` (review_pass prose unique); absence: `DELTA REVIEW`, `reviewBaseSha`, `preFixSha`, `maxCycles`, `cyclesRun`, `fixedInCycle`, `allCoverageGaps`, `for (let cycle` (skeleton guard), `review_loop`, `/review[- ]loop/i`
 - `reviewed: true`, `coverageGaps.length === 0`, `FAIL-FIXED`, `ALWAYS ready`, `Cheapest-sufficient validation`, `One build gate per phase`, `NEVER wrapped in`, `Gate 1 #2`, `gate1-final`, `No unauthorized GitHub side-effects`
@@ -275,7 +287,8 @@ In the SINGLE mode workflow's final Gate 1 (#2, `gate1-final` phase), retry atte
 - `src/assets/commands/dynamic-build.mds` — main build command source with inline SINGLE + WAVE workflow scripts
 - `dist/commands/dynamic-build.md` — compiled artifact pinned by test suite
 - `tests/build-mds.test.ts` — doctrine-literal pinning tests (sections 10, 12, 13)
-- `scripts/build-mds.ts` — unified MDS compiler (13 compiled hosts `ALL_HOSTS`; `DIST_FILES` = 14 including hand-authored `release.md` — SG-13 permanent divergence)
+- `scripts/build-mds.ts` — unified MDS compiler for both host kinds (command hosts → `dist/commands/`, generator hosts → `dist/agents/`); see the count-rule table above for what 13/14/14 each count. The pipeline itself — discovery, destination validation, the frontmatter strips, pruning — is documented in the `feature-knowledge-system` KB
+- `tests/fixtures/mds-manifest.ts` — shared name manifest for the suite: `MDS_COMMAND_HOSTS`, `MDS_GENERATOR_HOSTS` (`['git']`), `MDS_PARTIALS`, `HAND_AUTHORED_COMMAND_FILES`, `DIST_COMMAND_FILES`, `ALL_MDS_HOSTS` — tests derive counts from these instead of pinning literals
 
 ## Deliberate Exceptions (AC-0.4 gh-issue scope guard)
 

.devflow/features/external-model-routing/KNOWLEDGE.md

@@ -263,7 +263,7 @@ The old "override confirm" prompt is gone. The merge is purely additive (Devflow
 
 ## Mapping Engine (agent-models.json)
 
-`~/.devflow/agent-models.json` is a **deviations-only** mapping: agents that use their shipped defaults are omitted entirely. There is **no `previousModel` field** — shipped defaults are read live from `src/assets/agents/` source files at convergence time via `loadShippedDefaults()`.
+`~/.devflow/agent-models.json` is a **deviations-only** mapping: agents that use their shipped defaults are omitted entirely. There is **no `previousModel` field** — shipped defaults are read live at convergence time via `loadShippedDefaults()`, which walks `agentSourceDirs()` (`dist/agents/` then `src/assets/agents/`) first-wins. A registry agent that neither directory supplies is reported through the aggregate `onWarning` channel rather than silently defaulting.
 
 **Type precision**: `EFFORT_LEVELS` in `agent-models.ts` and `CLAUDE_MODEL_ALIASES` in `external-models.ts` are both `as const`, giving derived literal union types (`EffortLevel = 'low'|'medium'|'high'|'xhigh'|'max'`, `ClaudeModelAlias = 'haiku'|'sonnet'|'opus'|'fable'`). These flow through `AgentMapping.effort`, `EffectiveConfig.effort`, and `AgentRow.configuredEffort`/`originalEffort` as `EffortLevel` (not `string`). The one remaining `as EffortLevel` cast at the `readAgentMapping` parse site is sound: the `has()` check proves membership before the cast.
 
@@ -314,8 +314,10 @@ Effort is orthogonal to dormancy — it always applies regardless of proxy state
 ### `loadShippedDefaults` and `reapplyAgentMapping` — parallel execution
 
 Both use `Promise.all` for parallel I/O:
-- `loadShippedDefaults()` reads all agent `.md` files from `agentsDir()` concurrently.
-- `reapplyAgentMapping()` processes all agent files concurrently via `Promise.all` over the agent name list.
+- `loadShippedDefaults(dirs = agentSourceDirs(), opts?)` reads every agent `.md` in each directory of `agentSourceDirs()` concurrently (`readDirDefaults` per directory) and merges them **first-wins**, so `dist/agents/` supersedes `src/assets/agents/` for a name present in both. A missing directory on either side yields an empty map.
+- `reapplyAgentMapping()` processes all agent files concurrently via `Promise.all` over the agent name list. It accepts an optional `agentSourceDirs` (same convention, injectable for tests) and passes its own warning channel down to `loadShippedDefaults`; `revertExternalAgents` forwards both.
+
+**Registry-gap warning**: after the merge, `loadShippedDefaults` compares the resolved names against `getAllAgentNames()` and emits ONE aggregate `onWarning` message naming every registry agent no directory supplied, pointing at `npm run build:mds` (mirrors the installer's throw message, which fires on the same invariant). It does not throw — `devflow agents --list` must still render. This is the disclosure for a real silent failure: in a `build:cli`-only tree `dist/agents/` is absent and the generated agent has no `.md` source, so `resolveEffective` returns `model === undefined`, `reapplyAgentMapping` buckets the agent `'unchanged'`, and **disabling the proxy leaves a GPT-pinned agent unreverted** (PF-022). `devflow agents` passes `p.log.warn` as the channel; `reapplyAgentMapping` routes it into `ReapplyResult.warnings`.
 
 Warning collection is **deterministic**: each parallel task returns its local warnings alongside its bucket result; the outer loop aggregates in `allNamesList` insertion order. Warnings are emitted to `opts.onWarning` immediately for live feedback and also collected for the returned `ReapplyResult.warnings` array.
 
@@ -395,7 +397,7 @@ A user who hardened `settings.json` to `0600` (to protect `ANTHROPIC_API_KEY`) n
 - **Short-circuiting the disable settings pass with `||`**: `removeProxyHooks(s) || _stripProxyEnvFromObject(s, port)` leaves `ANTHROPIC_BASE_URL` set when hooks are present. Both operations must run unconditionally — see `applyDisableToSettings`.
 - **Running `reapplyAgentMapping` before proxy preflight completes**: preflight can force `proxyEnabled=false`, and the dormancy logic depends on the final resolved value. In init, the guard is placed immediately after the proxy preflight block.
 - **Calling `process.exit()` inside a finally-guarded scope in the TUI**: cleanup must be wired via Promise `resolve()`. Any `process.exit()` inside `finally` terminates without running cleanup and causes event-loop issues (avoids PF-014).
-- **Using previousModel in agent-models.json**: The mapping has no `previousModel` field. Shipped defaults are always read live from `agentsDir()` source files. Caching a previousModel creates stale drift when source agent files are updated.
+- **Using previousModel in agent-models.json**: The mapping has no `previousModel` field. Shipped defaults are always read live from the directories `agentSourceDirs()` names — `compiledAgentsDir()` first, `agentsDir()` as the fallback — so a generator host's compiled artifact is the default for the agents it produces. Caching a previousModel creates stale drift when those agent files are updated.
 - **Duplicating the dormancy predicate**: `isDormantExternalModel(model, proxyEnabled)` from `external-models.ts` is the single source of truth. Do not inline `!isClaudeModelName(model) && !proxyEnabled` at call sites.
 - **Pre-spawn doctor gating (chicken-and-egg)**: The relay's `doctor` subcommand probes the relay port to confirm it is running — a not-yet-started relay makes that probe fail (exit 1). A pre-spawn gate is therefore always unsatisfiable on a cold path and invisible to unit tests that mock doctor exit 0 (found during the first live enable). Doctor must gate post-spawn only, after the relay is confirmed up (D-EFR-2).
 - **D-EFR-3: Never mock the routing-runtime subprocess without a paired real-binary test**: any test that mocks the routing-runtime subprocess must be paired with at least one CI-executed test that does not. The specific trap (PF-016 reproduced exactly): `tests/integration/**` is excluded from `npm test` by `vitest.config.ts` while CI runs only `npm run build && npm test` — a real-binary test placed in `tests/integration/` would never execute in CI. Place real-binary tests in `tests/` (not `tests/integration/`).