In a multi-iteration tool loop, buildResumePrompt anchored on the last user
message and walked forward through preceding tool results. The only user
message stays at index 0, so each iteration re-sent the entire transcript over
stdin to the resumed CLI session, duplicating the original query plus a
growing stack of tool results into the on-disk session every turn. The agent
saw its own context bloating and called it "compaction." Anchor on the last
assistant message and slice forward instead — the genuine delta since the
last turn already on disk. Same bug+fix in droid-cli (copy of pi-claude-cli).
Quick chat also created a fresh pi/Claude CLI session per user message and
faked continuity by stuffing the last 50 chat_messages into the prompt as
"## Previous Conversation". Replace that with real session continuity:
chat_sessions gains a cliSessionFile column (migration 56); ChatManager
opens the existing pi SessionManager file when present and creates a fresh
one (persisting its path) on the first turn. The prompt now carries only
the new user content.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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
Merges FN-3049 to replace all runtime dynamic imports of `@fusion/engine` with static imports that esbuild can inline into the CLI bundle, preventing the known regression where the published `npm i -g @runfusion/fusion` silently failed at the import site. Also ships FN-3024, adding an experimental a
Fusion-Task-Id: FN-3049
This merge delivers v0.12.0, which includes a new experimental agent onboarding modal (`ExperimentalAgentOnboardingModal`) that streamlines the handoff from onboarding to the create-agent form, along with backend lifecycle tests and documentation for the experimental flag. The release bumps all work
Fusion-Task-Id: FN-3024
The merge consolidates mock helpers across packages by creating a new shared `mockCoreEngine.ts` in the CLI package and strengthening the core/engine helpers, then migrating both CLI command tests and the dashboard's `AgentsView` tests to use the canonical helpers. It also publishes SSE architecture
Fusion-Task-Id: FN-3053
Added a new `droid-cli-probe` module to the dashboard package with test coverage, implementing a CLI probe capability for the droid system.
Fusion-Task-Id: FN-2979
This merge implements a recoverable pending-login auth flow (FN-3047) across three phases: cancellable in-progress auth routes in the API, an extended auth client contract, and updated ModelOnboardingModal/SettingsModal UI with proper cancel-button visibility during logout. It also adds `dashboard-r
Fusion-Task-Id: FN-3047
This merge introduces canonical mock helper modules across the dashboard and engine packages, and migrates representative test suites to use them for consistency. The changes add four new mock helpers (`mockApi.ts`, `mockLucide.ts`, `mockCore.ts`, `mockCoreEngine.ts`) and harden the API mock proxy's
Fusion-Task-Id: FN-3044
Completes Step 2 of FN-3042 by scaling board typography styles across TaskCard and QuickEntryBox components, adding a new CustomModelDropdown stylesheet, and introducing keyboard focus-visible styles for theme selector controls. The merge includes new regression tests for board font scaling, resolve
Fusion-Task-Id: FN-3042
- Replace implicit theme swatch rendering with explicit sample definitions in the custom model dropdown styling
- Add ThemeSelectorSwatchContract regression tests to lock the swatch contract and prevent visual drift
- Restore workspace lint/typecheck gate behavior in onboarding-related settings flows
- Document the explicit swatch sample contract in dashboard README and related onboarding prompt/context updates
Fusion-Task-Id: FN-3043
- Add backend onboarding session service and legacy API routes for start, state, and import/export generation
- Introduce experimental AgentOnboardingModal UI with stream-state handling, model selector wiring, and onboarding launch flow in AgentsView
- Add onboarding gating via prompt overrides/store updates and include search behavior fix for partial-match filtering
- Expand dashboard/core tests for onboarding API, modal behavior, agents/settings integration, and prompt override handling
Fusion-Task-Id: FN-3025
Merges FN-3009 (remote settings migrated to global scope with updated schema and docs), FN-3034 (legacy routines compatibility layer: agentId backfill on load and schema validation for old format), and FN-3039 (Nerd Font glyph asset bundled into the terminal with CSS prioritization). Core settings s
Fusion-Task-Id: FN-3034
This merge migrates Fusion's settings architecture to a unified global scope (FN-3009), consolidating settings types, defaults, merge semantics, updater logic, and routing into a centralized system that aligns dashboard remote settings with global configuration. Secondary changes include adding the
Fusion-Task-Id: FN-3009
Exports the research settings resolver from `@fusion/core` types to fix dashboard alias builds, ensuring the resolver is available when imported through the dashboard's module alias configuration.
Fusion-Task-Id: FN-2995
The vitest child-process guard wrapped exec/execFile without preserving the
`[util.promisify.custom]` symbol, so awaited `execAsync` resolved to a raw
stdout string instead of `{stdout, stderr}`. That single regression cascaded
through ~60 "failing" tests across cli, core, engine, and dashboard whose
production code was actually correct. Also relax the AI-CLI blocklist for
cheap introspection (--version/--help/which …), give SIGTERM'd subprocesses a
brief grace period before being flagged as "left running", fix a few real
test-side bugs uncovered along the way (executor mock step transitions, iOS
last-resort keyboard path, mission SSE replay tests racing with the real AI
agent), and convert dashboard route tests' dynamic `await import("../server.js")`
to static imports so first-test timings drop from 2–5s to <200ms.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three coordinated fixes for the FN-2978 incident class — auto-requeues
that orphaned committed work and watchdog kills on long verification runs.
**Auto-requeue branch reuse** (executor.ts, worktree-pool.ts)
- executor.ts:1782 now uses `task.branch || fusion/<id>` so persisted
branches are honored on requeue. Previously the hardcoded fallback
always tried to re-create the original branch, hit a conflict with
the prior run's ref, and got suffix -2/-3. Other call sites already
honor task.branch — this aligns the worktree-acquisition path.
- worktree-pool.ts:181 prepareForTask now probes existing branches with
`git rev-parse --verify` and checks them out as-is. Falls through to
suffixed creation only when the branch is genuinely in use by another
live worktree. Previously force-reset with `checkout -B`, destroying
prior commits.
- New private reconcileStepsFromGitHistory walks `git log
baseCommitSha..HEAD` for `feat(FN-X): complete Step N` commits and
marks matching steps[] as done so resumes don't redo committed work.
**Manual reset endpoint + UI** (dashboard)
- POST /api/tasks/:id/reset (requires `confirm: true`) — clears worktree,
branch, all retry counters, resets steps[] to pending, moves to todo.
Distinct from /retry which is the soft-resume path.
- Reset button alongside Retry in TaskDetailModal with confirm dialog,
wired through useTasks → AppModals → API.
**fn_run_verification tool** (run-verification-tool.ts, executor.ts)
- New custom tool wrapping test/lint/build commands with a heartbeat
callback (per-line + 60s synthetic), 200KB head+tail output cap, hard
timeout with SIGTERM→SIGKILL escalation, and auto-bootstrap detection
for missing node_modules. Prevents the inactivity watchdog from
killing sessions during long compiles.
- Cross-platform via `shell: true` (Node picks /bin/sh on POSIX,
cmd.exe on Windows). Prompt section in EXECUTOR_SYSTEM_PROMPT and
EXECUTOR_PROMPT_TEXT instructs agents to prefer package-scoped
verification first and reserve workspace-scoped runs for final
integration.
**Tests** (64 passing)
- detect-pseudo-pause.test.ts (27 tests) — covers all 7 regex patterns,
structural fallback, FN-2978 regression text.
- reconcile-step-regex.test.ts (25 tests) — pins the commit-message
regex against a wide variant set.
- run-verification-command.test.ts (12 tests) — basic execution, output
capture, heartbeat callbacks, timeout, error handling. POSIX-specific
cases (multi-cmd `;`, `>&2`, `\$USER`) gated behind itPosix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The merge hardens the cloudflared installation backend with fallback logic and architecture detection, updates the SettingsModal UI with clearer manual installation guidance, and adds comprehensive test coverage for the remote access routes including the new memory-based settings registration flow.
Fusion-Task-Id: FN-2977
Merges external Tailscale funnel detection (FN-2976) — adds types, detection logic, status inclusion, kill flow, and a dedicated UI panel for funnel processes started outside Fusion — alongside custom AI providers API routes and a new settings UI section (FN-2965).
Fusion-Task-Id: FN-2976
Merged v0.9.4 bringing custom AI providers to the dashboard — SettingsModal now surfaces provider configuration with an onboarding disclosure flow, backed by new API routes (`/custom-providers`) and comprehensive test coverage. The engine adds self-healing with a ghost-review fallback recovery path,
Fusion-Task-Id: FN-2965
Hoist agentStore to outer scope so the finally block can close it
before removing the temp dir, preventing leaked SQLite handles when
the test errors mid-try.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Each agent now gets its own .fusion/agents/<id>/HEARTBEAT.md procedure
file instead of sharing a single project-wide file. A one-shot
migration in AgentStore.init() re-points existing agents off the legacy
shared path and copies the legacy file's contents into each agent's
new per-agent location so operator edits are preserved.
The HeartbeatTriggerScheduler now phase-aligns the first tick to
lastHeartbeatAt + intervalMs so a process restart resumes each agent's
existing schedule rather than waiting up to a full interval before
firing again. Overdue ticks fire promptly within a small jitter window
to avoid a thundering herd at boot.
Also fixes three pre-existing QuickChatFAB test failures introduced by
739e899b5: auto-select default model now switches to model mode whether
or not agents are present, the model tag only renders in model mode,
and one test scopes its option lookup to role="option" to disambiguate
the in-header tag from the dropdown entry.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The TUI splash/footer and the dashboard /api/health + /api/updates/check
each had their own walk-up resolver with different rules, so they could
report different versions and trigger spurious update-available banners.
Dashboard now uses a direct ancestor walk (matching the TUI's old logic)
and only falls back to the sibling-cli probe for dashboard-source dev
mode; logo.ts delegates to it so there is one source of truth.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds server-side validation for the `unavailableNodePolicy` field on the
PUT /api/settings route. Accepts only `block` or `fallback-local`;
rejects unknown strings and non-string values with 400.
- packages/dashboard/src/routes/register-settings-memory-routes.ts:
call `validateUnavailableNodePolicy` and throw `badRequest` on invalid input
- packages/dashboard/src/__tests__/routes.test.ts: cover accept,
invalid-value, and non-string cases
Fusion-Task-Id: FN-2952
When the tunnel was started before parseReadiness flipped to URL-required
(or by an external invocation), the manager's status URL stays null and
buildRemoteLoginUrlForTokenType used to fall back to a constructed
http://<configured-hostname>:<port> URL — which the user saw encoded
into the QR as e.g. "fusionstudio:4040".
Add a tailscale status --json fallback that recovers the public tailnet
DNS name and reconstructs the funnel URL (https://<machine>.<tailnet>.ts.net/).
If both sources fail, surface REMOTE_URL_NOT_READY rather than emitting a
non-public URL into the auth/QR payload.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Override the stored tailscale targetPort in the start handler using the
actual port the start request landed on (req.socket.localPort), instead
of trusting a value the browser supplied. window.location.port can be
wrong behind a reverse proxy, ssh forward, or split deployment, and
silently funneling the wrong process is worse than the prior
"port-not-configured" error.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Replace placeholder /remote/qr SVG (URL drawn as text) with real QR
rendered via the qrcode package; add format=terminal returning ASCII
QR for the TUI.
* Resolve the public tailscale funnel URL from captured CLI output
instead of constructing http://<hostname>:<port> from a configured
hostname label — that label was never used by `tailscale funnel` and
produced a non-public URL in the auth/QR link.
* Drop hostname requirement from engine + UI; only target port matters.
* Tighten tailscale parseReadiness to require a URL on the matched line
so the tunnel manager doesn't lock in `running` before the URL line.
* TUI: poll remote status, show ● tunnel indicator + URL in MainHeader,
bind Ctrl+Q to a global QR overlay (terminal ASCII), and switch the
in-Settings K shortcut to render the same ASCII QR.
* Auto-poll remote status in the dashboard while in `starting`/`stopping`
so the UI flips to running without reopening the modal.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add `triage` to VALID_TRANSITIONS["in-review"] so the dashboard's
`Request AI Revision` and `Rebuild Spec` actions work for in-review
tasks. moveTask now applies the same full reset on in-review → triage
as on in-review → todo (clears branch/baseBranch/baseCommitSha/summary/
recovery metadata and workflowStepResults) so the respec'd task starts
from scratch. The in-review task card's Move menu also gains Planning
as a destination.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When a worktree was rebased onto newer main, baseCommitSha remained an
ancestor of HEAD but the range baseCommitSha..HEAD then swept in upstream
main commits as if they were task changes (FN-2840 showed 33 files for a
4-file task). The previous display-recovery only fired when baseCommitSha
was no longer an ancestor, missing this case.
When enableDisplayRecovery is on and baseBranch is missing, also compute
merge-base(HEAD, main) and prefer it when it's a descendant of
baseCommitSha (tighter fork point). When merge-base is not a descendant
(FN-2855: baseCommitSha is on a deleted feature branch), keep
baseCommitSha so we don't widen to unrelated upstream files.
Also stabilizes the card-vs-Changes-tab divergence: resolution no longer
depends on whether a rebase happened between polls.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two fixes for the in-review/in-progress "files changed" count on task
cards and the Changes tab:
- Drop untracked files from /tasks/:id/diff and /tasks/:id/file-diffs.
At review time these are almost always build artifacts/cache/logs not
in .gitignore, not real task changes, and they inflated the count.
- Add display-only `enableDisplayRecovery` option to resolveDiffBase.
When a worktree was rebased onto origin/main after baseCommitSha was
recorded and baseBranch was not set, the prior code fell through to
HEAD~1 — undercounting to just the last commit's files (e.g. FN-2957
showed 2 in review, 6 after merge). Recovery now tries
merge-base(HEAD, main) / origin/main before HEAD~1.
Routes opt into recovery; the merger's mirrored copy
(resolveTaskDiffBaseRef) is intentionally untouched so merge-time
scope warnings still evaluate the strict task base.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Update dashboard mocks to push assistant text into session.state.messages
(matching real session.prompt() void return) and add a regression test for
array-shaped content. Add core unit tests for extractDreamProcessorResult
to lock in undefined/null safety.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
session.prompt() resolves to void; the reply lives in session.state.messages.
The dream route was awaiting its return value, so parsing crashed with
"Cannot read properties of undefined (reading 'match')" once the dream
completed. Extract the assistant text post-prompt and fail soft on
non-string input in extractDreamProcessorResult.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Self-review of the recovery cascade surfaced three issues; this commit
addresses all of them.
1. AI didn't actually receive the safety constraint under Layer 3.
The previous commit logged the safety preamble to the task log via
`store.logEntry`, but the merge agent doesn't read task log entries as
prompt context — so the AI was running blind. The "no silent
re-introduction of main's deletions" guarantee was therefore relying
*entirely* on the deterministic verification gate (test + build),
which is correct as a backstop but doesn't help the AI produce a
correct first attempt.
Fixed by threading `preMergeRebaseFallthrough` through
`MergeAttemptParams` → `executeMergeAttempt` → `runAiAgentForCommit` →
`MergePromptParams` → `buildMergePrompt`, where it now injects an
explicit "⚠️ Pre-merge rebase recovery exhausted" preamble at the top
of the user prompt with three concrete rules:
- Prefer main's deletion when branch re-adds removed lines
- Prefer main's version on ambiguous hunks
- Call `fn_report_build_failure` rather than commit a regression
Also includes the original rebase failure message (truncated) so the
AI has diagnostic context.
The truncated-context retry path also forwards the preamble — it's the
safety constraint, not bulk context, so we keep it even when stripping
diff stat / commit log to fit the window.
2. Layer 2's branch-restore could fail with "uncommitted changes".
When a cherry-pick midway through Layer 2's replay fails, the worktree
is in a half-applied state with conflicts in the index. The previous
restore did `git checkout <branch>` (no -f) followed by
`git reset --hard <originalSha>`. The plain checkout would refuse with
"would overwrite local changes" if there were unmerged paths,
preventing the reset from running and leaving the branch at the
half-replayed tip.
Fixed by reordering: hard-reset to the captured original SHA first
(this clears index/working tree of any cherry-pick state), then
`git checkout -f <branch>` to ensure HEAD points at the named branch,
then a final hard-reset to the original SHA as belt-and-suspenders.
Worst case the worktree is at the original branch tip — never worse
than where Layer 2 started.
3. Pre-existing unrelated lint error blocking workspace lint.
`packages/dashboard/src/server.ts` had an unused `resolve` import from
`node:path` left behind by a recent refactor that extracted
`PACKAGE_VERSION` into its own file. The user explicitly asked to
clean it up so workspace lint passes. One-line drop.
Tests + checks:
- Engine: 2886 / 2886 pass (added safety preamble didn't break any
existing prompt-content assertions)
- Core: 3120 / 3120 pass
- Workspace lint: clean
- Engine typecheck: clean
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
