## What happened
FN-8004's implementation work finished and passed review. The auto-merge
then failed with `Grok ACP turn failed: Internal error` — a ~20 second
provider blip — and the task was parked `status: "failed"` with 8 files
of complete, reviewed work stranded on its branch.
The park is the interesting part: `status: "failed"` is precisely what
tells recovery to stop. So a misclassification here isn't a missed
retry, it's **terminal**. Both recovery paths were disabled by the same
wrong verdict:
- `maybeRetryTransientMerge` (inline, 3 retries w/ backoff) — never
fired once (`mergeTransientRetryCount: 0`).
- `recoverTransientMergeFailures` (self-healing sweep, exists exactly to
rescue parked in-review tasks) — skipped it, gated on the same
classifier.
## Three defects fixed
**1. No AI-provider failure class existed.** The AI merge drives a real
LLM turn, but `classifyTransientMergeError` only modeled git/lease/spawn
faults. Adds `ai-provider-turn-failure`.
**2. ACP dropped the error detail.** `promptAcpSession` rethrew the SDK
error unchanged, discarding the JSON-RPC `code`/`data` — the only
evidence the fault was provider-side. ("Internal error" is just the
standard text for `-32603`.) It now preserves them, keeping the original
as `cause`:
```
Internal error (acp rpc code -32603, retryable)
```
Classification anchors on that envelope, **not** on the bare `"Internal
error"` — matching that unanchored would disguise genuine application
defects as retryable blips. Only provider-fault codes (`-32603`,
`-32000`..`-32003`) are retryable; caller-fault codes
(`-32600`..`-32602`) stay permanent, since retrying just repeats the
failing call.
**3. Sweep/inline asymmetry** (found while tracing; latent and
unreported). The inline gate accepted `isTransientError(msg) ||
classify(msg)`, but the sweep consulted **only** the classifier. So
`ECONNRESET` / `socket hang up` during a merge earned inline retries and
then went **invisible to the sweep** once parked — stranded forever. The
classifier now delegates to `isTransientError`, so both gates agree by
construction.
To keep that delegation from importing the detector's
`usage-limit-detector → logger` chain (the chain FN-5627 split the
classifier out to avoid, which would break
`notification-service.test.ts`'s partial `vi.mock`), the pure predicates
moved to the import-free leaf `transient-error-patterns.ts`, re-exported
from `transient-error-detector.ts`. All 13 exports preserved, verified
programmatically.
## Loosened budgets
Per request, so more self-heals. Both apply **only** to errors already
proven transient; the ceiling and
`merger:transient-failure-budget-exhausted` audit path remain.
| Budget | Before | After |
|---|---|---|
| `MAX_AUTO_MERGE_TRANSIENT_RETRIES` | 3 | 5 (backoff
5s/10s/20s/40s/80s) |
| `MAX_TRANSIENT_MERGE_RECOVERIES` | 2 | 5 |
The bump broke two suites that had hardcoded the old `3`. Rather than
swap in another magic number, both now derive the cap from the constant
so future tuning doesn't re-break them.
## Verification
- `pnpm test:gate` green · `pnpm lint` clean · engine + ACP typecheck
clean · `pnpm verify:fast` PASS (5/5)
- ACP plugin 230 tests green · Grok plugin 64 green · engine
transient/merge suites 136 green
- Regression tests assert the **invariant across every surface** (per
*Fix the Invariant, Not the Repro*), not just the reported Grok string:
both ACP runtime prefixes, all retryable/non-retryable rpc codes, both
SDK error shapes, network delegation, class-ordering, and negative cases
proving bare `"Internal error"` and real defects stay permanent.
- A test caught a genuine bug in my own code mid-review (nested-shape
message shadowing), now fixed.
- `notifier.test.ts > "awaiting approval"` fails — **confirmed
pre-existing on clean main**, unrelated.
## Note
FN-8004's own branch (`fusion/fn-8004`) is still unmerged and its work
looks complete. Once this lands, its merge should be retried separately.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Grok advertises promptCapabilities.image=false and ignores ACP image
ContentBlocks (live probe: NO_IMAGE). Path-based vision works when the
agent is given an absolute file path. Include path hints in chat prompts
from .fusion/chat-attachments and carry path on ChatImageContent for
file:// uris.
AcpRuntimeAdapter.promptWithFallback ignored options, so dashboard chat
images never became ACP ContentBlock image entries. Extract images from
prompt options and pass them through buildPromptBlocks for both acp-runtime
and the Grok vendored client.
Grok emits `_x.ai/session_notification` / `_x.ai/session/update` for
hook_execution status. The ACP SDK routes those to Client.extNotification;
without it, every successful post_tool_use hook logged -32601 Method not
found. Implement no-op extMethod/extNotification on default and bridging
handlers in acp-runtime and the Grok vendored copy.
Replace one-shot grok -p JSON with native grok agent stdio (ACP) for realtime
streaming, tool visibility, and multi-turn sessions. Vendor the ACP client
into fusion-plugin-grok-runtime, forward Fusion fn_* tools and operator MCP,
stage Fusion skills via --plugin-dir, authenticate per xAI headless docs, and
align project chat manager store resolution so Grok chat sessions can send.
- Greptile P2: `acp.active` now reflects the ACTUAL dispatch determinant
(FUSION_CLAUDE_ACP, which includes the operator force-override), not the
experimental flag alone — so the status isn't misleading when forced on/off.
- CodeRabbit/Greptile P2: add FNXC:ClaudeAcp comments to the new code blocks
per the AGENTS.md greppable-comment convention.
Already fixed in the prior commit (daa37d08c): the P1 "sticky env" / latch
(applyClaudeAcpEnable now recomputes each call + FUSION_CLAUDE_ACP_FORCE
override) and the enable->disable-on-same-env regression test.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two-reviewer pass (security + architecture) on KTD10 + the full Route A
increment: no code defects, no P0, merge-ready as a dormant increment. Applying
the P1 follow-ups:
- Add the feature changeset (@runfusion/fusion minor) — the one convention gap.
- KTD10 tests: fail-closed (bridge not resolved -> env stays unset -> -p) and
idempotency (second onLoad keeps the first published path).
- Document the two intentional, parallel MCP-forwarding paths (U10 engine-adapter
vs U11 provider-driver) so nobody double-forwards, and the known
ACP-path-token-usage=0 residual so U12 doesn't treat it as a bug.
Reviewers confirmed: dormancy invariant holds end-to-end (nothing sets
FUSION_CLAUDE_ACP=1; both flag+path required; -p is the default); OAuth pi path
untouched. 206/206 plugin tests, 333/333 pi-claude-cli tests, typecheck clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The acp-runtime plugin's onLoad now publishes the identity-pinned bundled
claude-code-cli-acp path to FUSION_CLAUDE_ACP_BRIDGE (when unset), so the
pi-claude-cli kill-switch resolves the bridge WITHOUT a manual env var — no
engine->plugin static coupling. Publishes the path only; the ACP transport stays
OFF until an operator sets FUSION_CLAUDE_ACP=1 (rollout gate). Explicit env
override wins; resolver is pinned to the plugin's node_modules/.bin shim.
204/204 plugin tests green (3 new KTD10 tests); typecheck clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Route A unblock + the first Route A code increment.
- U9 verdict recorded (plan OQ1 + docs/acp-contract.md): in an authenticated
interactive session the pinned claude-code-cli-acp 0.1.1 bridge forwards
session/new mcpServers to Claude, Claude invokes the forwarded Fusion tool,
and the call traverses the ACP permission gate (session/request_permission).
Both security-critical answers resolve positively — overturns the headless
NOT-GO chain (FN-6466/6467/6473/6476), whose only blocker was running
detached from the login keychain session.
- R17 (daemon auth) recorded and closed for the supported setup: creds are
macOS Keychain-only; the user's login-session fn daemon has keychain access
(the existing claude -p provider authenticates there), so the bridge does too.
- U10: thread an optional mcpServers list through the ACP runtime contract.
newAcpSession now forwards it (was hardcoded []); AgentRuntimeOptions (engine
+ plugin-local copy) gains the field; defaults to [] to preserve Route B's
read-only ask posture. Tool calls still route through the U5 permission floor.
Plugin typechecks clean; provider-session tests 12/12 (incl. 2 new forwarding
tests). U11-U13 (provider transport swap, picker/auth, workflow verify) remain.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
