This merge lands v0.12.0 with two major features: a droid-cli provider integration adding auth routes, status endpoints, and a settings toggle hook for controlling CLI-based authentication, plus a new experimental agent onboarding modal with a create-agent form. The release also stabilizes engine st
Fusion-Task-Id: FN-2981
- Remove ToolSearch prerequisite so custom MCP tools can be used during triage and plan generation flows
- Align built-in tool sets between provider wiring and prompt builder handling, including custom ls behavior
- Expand pi-claude-cli tests for event bridge, MCP config, prompt builder, and tool mapping regressions
- Add FN-2703 changeset and delivery documentation for the published @runfusion/fusion package
Triage with claude-sonnet-4-6 via pi-claude-cli kept looping on
fn_review_spec calls that were rejected by pi's validator with
"root: must be object". Parameterless MCP tools (schema
{type:"object", properties:{}}) emit zero input_json_delta events,
so partialJson stayed "" and the catch fell through to
finalArgs = "" — a string, which TypeBox's Type.Object({}) rightly
refuses. Default empty partialJson to {} so the call lands.
Also:
- Add a 2-step reminder loop in triage before swapping to the
fallback planning model — primary models that wrote PROMPT.md
but forgot fn_review_spec recover from a nudge, no need to pay
the cold-start tax of a new triage on a different model.
- Inject @runfusion/fusion's own pi extension into dashboard/
daemon/serve sessions and propagate the path to createFnAgent
via setHostExtensionPaths so fn_* tools register globally
without requiring `pi install npm:@runfusion/fusion`.
- Drop the "historical" qualifier from replayed tool labels —
Claude was reading "TOOL RESULT (historical Read):" as
"previous session, ignore" and looping on verification.
- Remove subprocess-lifecycle stderr debug logs that landed for
hang diagnosis — root cause is fixed, the noise can go.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Root cause of "fn_review_spec was never called" tracked through the diagnostic
chain: the prompt actually sent to triage agents has zero `fn_*` tokens —
because resolveAgentPrompt("triage", ...) returns the BUILTIN_AGENT_PROMPTS
default-triage template (TRIAGE_PROMPT_TEXT in core/agent-prompts.ts), and that
template was forked from an older version that never had the "MUST call
fn_review_spec()" workflow nor any fn_-prefixed tool names. The fallback
`|| TRIAGE_SYSTEM_PROMPT` in engine/triage.ts never fires because the core
template is non-empty.
So the model writes PROMPT.md, doesn't see any instruction to review it, and
ends. zai/glm-5.1 happened to call fn_review_spec from training-pattern
inertia; Sonnet via pi-claude-cli stopped at write — same prompt, same bug.
Replace TRIAGE_PROMPT_TEXT with the engine's up-to-date TRIAGE_SYSTEM_PROMPT
verbatim (fn_-prefixed tools, fn_review_spec workflow, subtask breakdown,
project-commands handling, frontend UX criteria injection). Also remove the
diagnostic-only console.error lines added during this debugging session — the
core fix is now elsewhere and the noise isn't worth keeping.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User reports "no custom tool refs to rewrite (tools=18)" — meaning Context.tools
is populated but the prompt body doesn't contain any of the names. Need the
first 200 chars of the prompt and a sample of custom-tool names to figure out
whether the engine's customPrompt is reaching us at all or whether names
differ from what we expect.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds stderr lines so we can verify the rewrite is actually firing on a given
session ("rewrote N custom tool ref(s) [fn_review_spec×3, ...]") vs. silently
no-opping (no tools, no matches). Helps distinguish "fix not deployed yet"
from "fix not effective" without redeploying instrumentation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Triage system prompts read "MUST call fn_review_spec()" but Sonnet 4.6 routed
through pi-claude-cli writes PROMPT.md and silently skips the call — even with
the addendum explaining the deferred-tool protocol. Confirmed by the FN-2564
agent log: model called other MCP tools (fn_memory_search, fn_task_list) fine
but consistently never reached fn_review_spec, leaving triage looping on
"fn_review_spec was never called" and falling back to zai/glm-5.1 every time.
Rewrite bare `fn_*` (and any non-built-in custom tool name) references in the
system prompt to their `mcp__custom-tools__fn_*` form before sending. The
prompt now literally says "call mcp__custom-tools__fn_review_spec()" so the
model has no inference step, and the deferred-tool reminder Claude Code injects
matches verbatim. Word-boundary safe (won't touch fn_review_specifier) and
idempotent (won't double-prefix already-MCP-named occurrences).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Claude Code 2.x defers MCP tool schemas behind ToolSearch — tools are listed
but calling them by short name (e.g. fn_review_spec) returns "tool not found"
until ToolSearch loads the schema. Triage system prompts that say "call
fn_review_spec()" leave the model guessing whether to use the bare name or
the mcp__custom-tools__ form.
Append a system-prompt section listing every custom tool with its short and
MCP-prefixed names plus the ToolSearch + call protocol. Drives off the
Context.tools we now already thread through, so plain chats with no custom
tools get no addendum.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- TaskCard: four catch((err: any) => err.message) promise handlers in
archive/unarchive/delete/move → catch((err) => getErrorMessage(err)).
- InlineCreateCard + QuickEntryBox: .catch((err: any)) model-load handlers
→ getErrorMessage(err) with existing @fusion/core import.
- TerminalModal: drop (navigator as any).maxTouchPoints — modern lib.dom
types already expose the property.
- serve.ts: remove unused any annotation on OpenRouter model mapper; the
array element type is already inferred from json.data.
- pi.js, runtime-resolution.ts, dashboard.ts, serve.ts, dev-server-port-
detect.ts, devserver-manager.ts: drop now-stale eslint-disable comments
that the cleanup made redundant.
Fix a prompt-builder regression surfaced by agent's `any` cleanup: toolCall
with a raw string `arguments` field must be preserved verbatim (JSON-quoted)
rather than coerced to `{}`; restores a previously-passing test.
Then promote @typescript-eslint/no-explicit-any from warn → error. Future
new anys must either come with a one-line disable + justification or use a
real type. Workspace is now lint-clean (0 problems).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Parallel subagent pass: four typescript-pro agents on non-overlapping scopes.
Patterns applied:
- catch (err: any) { ... err.message ... } → catch (err) { ... getErrorMessage(err) ... }
using the new @fusion/core helper. Bare catch {} where the error was unused.
- SQLite row types: defined typed XxxRow interfaces per table and cast
.all()/.get() results via `as unknown as XxxRow[]` (the double cast is
required because better-sqlite3 returns Record<string, SQLOutputValue>).
- rowToX(row: any) converters: typed argument with the matching row interface.
- Dynamic settings key writes: (settings as Record<string, unknown>)[key].
- React event handlers and setState callbacks: inferred types or concrete
React.{Mouse,Change,Form}Event<...> where needed.
- pi-claude-cli: local PiMessage / PiContext duck types to avoid re-typing
pi-ai concrete shapes; typed Claude stream event message fields.
72 files changed, ~400 anys eliminated. Typecheck passes across the workspace.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the stray useClaudeCli settings checkbox + onboarding question
with a proper provider-card UX. The card lives next to OAuth + API-key
cards in onboarding and settings, with Enable/Disable + Test actions.
Backend:
- Vendors rchern/pi-claude-cli@0.3.1 as packages/pi-claude-cli
(MIT, attribution in UPSTREAM.md). Lets us bump peer-dep on
pi-coding-agent in lockstep with Fusion (upstream pinned ^0.52.0
vs ours ^0.62.0) and fix bugs without waiting on upstream.
- Adds @fusion/pi-claude-cli as a workspace dep of @runfusion/fusion
so users don't have to `npm install -g pi-claude-cli` manually.
- serve/daemon/dashboard conditionally load the extension via
discoverAndLoadExtensions() when GlobalSettings.useClaudeCli is on;
no side-effects on user ~/.fusion/agent/settings.json.
- New GET /api/providers/claude-cli/status: claude --version probe
+ toggle state + cached extension resolution.
- New POST /api/auth/claude-cli: flips useClaudeCli, refuses if the
claude binary is missing, fires the existing skill-backfill hook.
- /api/auth/status now injects a synthetic {id:"claude-cli", type:"cli"}
provider entry so onboarding + settings see a consistent list.
Frontend:
- New ClaudeCliProviderCard component shared between ModelOnboardingModal
and SettingsModal's Authentication section.
- New AuthProvider.type = "cli" variant.
- Removed the old "Route AI calls through the Claude CLI" checkbox from
Global Models settings and the opt-in step from the onboarding wizard.
- ProviderIcon gets a composite Anthropic-mark-plus-terminal glyph for
the claude-cli provider id.
Tests:
- 8 unit tests for extension resolution (@fusion/pi-claude-cli is
workspace-linked so these run in-tree).
- 2 unit tests for the binary probe.
- Existing /auth/status tests filter out the new synthetic entry so
they keep asserting structural OAuth/API-key behavior in isolation.
- The vendored package's own 296 tests still pass unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
