mirror of
https://github.com/NanmiCoder/cc-haha
synced 2026-07-27 15:13:37 +08:00
1401107509
* feat(agent-tool): rewrite contradictory review verdicts at the harness layer
The code-reviewer and security-reviewer subagents end with a parseable
verdict (`REVIEW: APPROVE` / `SECURITY: PASS`) and their prompts forbid
emitting the positive form alongside `[CRITICAL]` or `[HIGH]` findings.
That rule is enforced only in prose, so when the model slips, the
parent agent reads "APPROVE" and ships.
Add a pure post-processing pass over the subagent's final text in
`finalizeAgentTool`: detect the verdict line, look for severity tags
above it, rewrite to `CHANGES_NEEDED` when they conflict, and prepend
a one-line audit notice so the parent (and a human reading the
transcript) can see the harness intervened. Logs a
`tengu_agent_sentinel_mismatch` analytics event for fleet visibility.
Debugger's `ROOT CAUSE: FOUND` is intentionally out of scope — its
"evidence from a real command" rule isn't mechanically verifiable.
Tested: bun test src/tools/AgentTool/sentinelCheck.test.ts (11 cases)
Confidence: high
Scope-risk: narrow
* feat(agent-tool): show subagent_type in tool-use transcript line
Each `Agent` tool call previously rendered as just its description in
the transcript, which made it invisible whether the main agent picked
a specialist or fell back to the default. Diagnosing "why didn't it
use code-reviewer?" required reading analytics or guessing.
Append `→ <subagent_type>` to the rendered description when the type
is explicit. Omitted (general-purpose default or fork) keeps the old
format — absence of the marker is itself the signal.
Tested: visual smoke only — pure UI render change with no behaviour
Confidence: high
Scope-risk: narrow
* feat(agent-tool): nudge verification subagent after edits accumulate
Verification is the only built-in agent that's *required* by the
session-specific guidance for non-trivial implementations, but the
trigger condition ("3+ file edits / backend / infra change") is
evaluated by the main agent from its own work — a classic LLM
self-audit failure mode. The model can convince itself that 4 edits
"are mostly one core change with small adjustments" and skip
verification entirely.
Add a `verification_gate_reminder` attachment that fires once when
the main thread accumulates `editCount >= threshold` (default 3)
file-mutating tool uses (Edit / Write / NotebookEdit) without
invoking the verification subagent. The reminder renders as a
`<system-reminder>` block on the next turn telling the model to
spawn `subagent_type="verification"` before reporting completion,
along with a permission to skip if the change is genuinely trivial.
A verification subagent invocation resets the counter; a previously
fired reminder suppresses re-firing until the next reset, so the
nudge isn't spammed every turn.
Subagents don't see the reminder (it's main-thread only). The gate
also no-ops when the verification agent isn't loaded in the current
session (e.g. SDK build with built-ins disabled). Disabled by
`CLAUDE_CODE_VERIFICATION_GATE_OFF=1`; threshold tuned via
`CLAUDE_CODE_VERIFICATION_GATE_THRESHOLD=N`.
Tested: bun test src/utils/verificationGate.test.ts (10 cases)
Confidence: high
Scope-risk: moderate
* feat(agent-tool): tighten subagent routing and ship real coordinator mode
The flat 15-agent routing has three known weak spots that this commit
addresses together because they share AgentTool.tsx call-path edits:
1. **General-purpose default is a vacuum cleaner.** Omitting
`subagent_type` falls back to general-purpose, which has full tool
access and cheerfully takes on review/security/perf/migration work
that a specialist would have done better and cheaper. Add a regex
heuristic over the prompt: when it strongly signals a specialist
(e.g. "code review", "security audit", "root cause", "migrate from
X to Y") and that specialist is available, refuse the default and
throw an error pointing the model at the right type. Disabled with
`CLAUDE_CODE_GP_DEFAULT_STRICT=0`. Skipped in fork and coordinator
modes (they have their own routing).
2. **No bound on verifier-style loops.** The verification prompt says
"On FAIL: fix, resume, repeat until PASS" — if the verifier
incorrectly keeps flagging the same change, the loop only stops
when the token budget runs out. Add a per-session counter for each
built-in agent type. Cap defaults are 5 for verification (loop
target) and 8 for everything else. When exceeded, throw with a
message asking the model to consult the user. Caps tunable per
type via `CLAUDE_CODE_AGENT_LIMIT_<TYPE>=N`; off via
`CLAUDE_CODE_AGENT_LIMITER_OFF=1`. Counts main-thread invocations
only.
3. **Coordinator mode was a stub.** The `coordinator/workerAgent.ts`
that `getBuiltInAgents()` requires when `COORDINATOR_MODE` is on
was an auto-generated Proxy returning nothing useful. Replace with
a real registry: a new `WORKER_AGENT` (full-tool-access generic
delegate matching the existing coordinator prompt's
`subagent_type: "worker"` references) plus the 11 specialists
minus general-purpose, claude-code-guide, statusline-setup. Also
harden the coordinator's own prompt with a "you orchestrate, you
do not implement" rule (no direct Edit/Write/NotebookEdit) and a
specialist roster table so the coordinator picks the right
delegate. Improve the AgentTool error message when coordinator
mode receives a request without `subagent_type`.
Constraint: cannot restrict the coordinator's own tool pool at the
tool layer without invasive changes to main-thread tool assembly —
enforcement is prompt-level, consistent with the rest of the
architecture.
Rejected: a tighter heuristic that requires a domain-specific phrase
(e.g. "review THIS PR") was considered for #1 but rejected — too
narrow to catch the routing problems we actually see in fleet logs.
Tested: bun test src/tools/AgentTool/specialistRouter.test.ts
bun test src/tools/AgentTool/invocationLimiter.test.ts
bun test src/coordinator/workerAgent.test.ts
(29 cases total, all pass)
Confidence: medium
Scope-risk: moderate
Directive: if specialistRouter false-positives in production, prefer
tightening the regexes over removing the gate — the failure mode
it catches (general-purpose taking specialist work) is worse than
the failure mode of an extra retry.
* docs(desktop): add local Chrome DevTools MCP testing guide
Document the workflow for driving the desktop UI end-to-end via
Chrome DevTools MCP without going through Electron. The pure-browser
dev path normally fails — CSP blocks non-loopback origins, the
H5-disabled CORS gate blocks cross-origin browsers, and the desktop's
loopback exemption clears the H5 token so even the loopback path
401s. The stable bypass is a Vite proxy that same-origins API
traffic to the renderer.
This commit covers the docs side only:
- `docs/desktop/10-local-mcp-testing.md`: full how-to with Quick
Start, MCP control snippets, can/can't test matrix, the exact
reproduction recipes for each `feat/agent-routing-hardening`
change (sentinel, gp-strict, limiter, verification gate, routing
observability), a Windows-specific `electron:dev` workaround, and
a fault dictionary mapping the symptoms each blocker produces back
to its cause.
- `docs/desktop/index.md`: link the new section in the docs index.
- `.kiro/steering/desktop-mcp-testing.md`: concise rule file with
`inclusion: manual` so it loads via #-mention only when relevant,
not as default-included context.
The Vite proxy itself is intentionally NOT committed here. The docs
explain it as a dev-only change the developer adds locally and
either reverts or commits to a separate `feat/desktop-browser-dev-proxy`
branch — the production renderer relies on Electron IPC for the
server URL and must not see the dev proxy in main.
Tested: docs render via VitePress local preview; steering file
loads via Kiro `#desktop-mcp-testing` mention; instructions
verified against the live setup that produced
artifacts/desktop-bootstrapped.png and artifacts/desktop-agents-page.png.
Confidence: high
Scope-risk: narrow
* feat(desktop): unblock pure-browser dev MCP smoke with proxy + forceH5
The standalone-browser dev workflow (vite + Chrome DevTools MCP)
fails three ways without these escape hatches:
1. CSP `connect-src` only allows loopback origins, so the renderer
can't talk to the API server cross-origin even on a LAN.
2. With H5 access disabled the server returns 403 to all
cross-origin browser requests.
3. With H5 access enabled the desktop's loopback exemption
(`requiresH5AuthForServerUrl` returns false for `localhost:1420`)
calls `setAuthToken(null)`, so `buildSessionWebSocketUrl` omits
`?token=`, and the session WS handshake 401s into a 1006 close
reconnect loop. The UI shows "处理中..." but the message never
reaches the server — `wsManager.connections` stays `{}`, server
stdout is silent, and no `POST /api/sessions/.../message` ever
fires.
Two dev-only changes break that deadlock without affecting
production:
- `desktop/vite.config.ts`: add `server.proxy` for `/health`,
`/api`, `/ws`, `/local-file`, `/preview-fs` to `127.0.0.1:3456`.
Same-origins all backend traffic so CSP and CORS pass. Vite dev
only — Electron production path is unaffected because Electron
resolves the server URL via IPC at runtime.
- `desktop/src/lib/desktopRuntime.ts`: add a `?forceH5=1` query
param check in `initializeBrowserServerUrl`. When present, treat
the URL as needing H5 auth even if it's loopback, so the H5 token
survives bootstrap and gets attached to the session WS URL. In
production the param is never set, so behavior is identical.
Workflow with both in place:
http://localhost:1420/?serverUrl=http%3A%2F%2Flocalhost%3A1420
&forceH5=1&h5Token=<TOKEN>
Verified end-to-end via WS sniffer + server log:
- ws://localhost:1420/ws/<sid>?token=... opens
- recv {"type":"connected","sessionId":"..."}
- server logs "[WS] Client connected" + "Starting CLI for ..."
- Sending a "use Task with subagent_type=Explore" prompt produced
a `tool_use_complete` for the Agent tool with
`input.subagent_type="Explore"` exactly as routed.
Also rewrites `docs/desktop/10-local-mcp-testing.md` to record:
- both patches as required steps (was missing the forceH5 piece)
- exact PowerShell commands to mint a one-shot H5 token
- the WebSocket sniffer script and three-signal diagnosis recipe
for "is the backend actually processing"
- the Item 5 (routing observability) gap: the change in
`src/tools/AgentTool/UI.tsx` only affects CLI/Ink rendering;
the desktop's separate `ToolCallGroup.tsx` doesn't surface
`subagent_type` and would need its own follow-up patch.
Tested: ran `bun run src/server/index.ts` + `bun run dev` in the
desktop/, opened the documented URL via Chrome DevTools MCP,
sent a Task-with-explicit-subagent_type chat message, observed
the full WS frame sequence and server log lines listed above.
Confidence: high
Scope-risk: narrow
Directive: do not generalize the forceH5 escape hatch to a
permanent feature — when it's needed, the dev-mode
loopback-clears-token branch is doing the right thing for
Electron and should not be removed.
* feat(desktop): show subagent_type badge next to Agent tool header
The CLI/Ink change in src/tools/AgentTool/UI.tsx surfaces routing
in the terminal transcript as `<description> → <subagent_type>`,
but the desktop renderer is independent (`ToolCallGroup.tsx` /
`AgentCallCard`) and was dropping that signal entirely. Diagnosing
"why didn't it use code-reviewer?" required reading analytics or
guessing.
Add a small `→ <subagent_type>` chip next to the Agent header in
AgentCallCard. The badge mirrors the CLI behavior:
- Renders only when `input.subagent_type` is a non-empty string;
general-purpose / fork / omitted defaults stay quiet.
- Carries `title="subagent_type: <name>"` for hover detail.
- Uses the same border-pill treatment as other secondary
transcript chips so it doesn't compete with the description.
Verified end-to-end via the local Chrome DevTools MCP path: a
real chat with `subagent_type=Explore` renders `Agent → Explore
查找specialistRouter相关文件` in the live transcript, screenshot
saved to artifacts/desktop-routing-arrow-highlighted.png.
Tested: cd desktop && bun run test -- --run -t "subagent_type"
2 tests pass:
- renders subagent_type as → Explore badge next to Agent header
- omits subagent_type badge when input has no subagent_type
Confidence: high
Scope-risk: narrow
* feat(desktop): MCP tool visibility, per-tool toggle, marketplace install
Surface MCP capabilities so the agent never under-uses a configured server,
and add a one-click install path for new servers.
Tool visibility & control:
- GET /api/mcp/:name/tools returns the live tool catalog (handles disabled /
needs-auth / host-preflight-failed / connected states)
- POST /api/mcp/:name/tools/:tool/toggle hides a single tool from the agent,
persisted to ~/.claude.json disabledMcpTools (user scope, all projects);
fetchToolsForClient filters on it and the cache is invalidated on toggle
- Desktop MCP detail page gains an Overview/Tools tab strip (both edit and
read-only views) listing name, description, annotations, JSON schema, plus
an inline enable/disable switch with a "hidden from agent" marker
Marketplace:
- Curated builtin catalog (14 servers across 8 categories) plus opt-in remote
catalog sources cached under ~/.claude/cc-haha/mcp-marketplace.json
- Endpoints: list / refresh / add-source / remove-source
- Desktop marketplace page with category grouping, install dialog (with
required-env prompts), and source management
Plugins:
- Plugin list rows get inline enable/disable + uninstall, matching the MCP
settings UX, instead of requiring users to open the detail panel
- Extract shared ToggleSwitch component
Tested: bun test src/server/__tests__/mcp.test.ts (tools list + per-tool
toggle + marketplace), cd desktop && bun run lint, mcpSettings vitest, and a
browser smoke pass via the local Chrome DevTools MCP flow (tool toggle
persistence, marketplace render, plugin enable/disable + uninstall dialog).
Not-tested: full bun run verify gate
Scope-risk: moderate
* fix(agent-tool): rephrase gp-strict error to stop poisoning sessions
Live testing surfaced an unintended side effect of the gp-strict
redirect message. The original wording embedded attribute-style
fragments like `subagent_type="code-reviewer"`. When that error
came back to the model as a tool result, the model copied the
shape into a TEXTUAL `<tool_use name="Agent">{...}` block instead
of issuing a real tool call — and every subsequent Agent call in
that session degraded to text. Confirmed by inspecting the
session JSONL via /api/sessions: messages 4..N in the affected
session were `assistant: text(...)` carrying simulated tool_use
XML, not `tool_use` content blocks. Reproduced with two unrelated
models (claude-haiku-4.5 and mimo-v2.5-pro) on the same poisoned
session, and crucially DID NOT reproduce in a fresh session with
either model — so the failure is the message text inducing
text-form mimicry, not the model itself.
Fix:
- Extract the redirect text into `formatSpecialistRedirectMessage`
in specialistRouter.ts so it has its own test surface.
- Rephrase as prose that names the parameter and value without
any `key="value"` assignment-and-quotes form. New shape:
"This task looks like a job for the code-reviewer specialist
rather than the general-purpose default. Re-call Agent with
the subagent_type parameter set to code-reviewer. ..."
- Wire AgentTool.tsx through the helper so the throw site is one
line.
Add a regression assertion to specialistRouter.test.ts that the
formatted message contains NO `key="value"` shapes, no XML-ish
open tags, and no `"subagent_type":` JSON fragments — across all
ten specialist names. This is the property the live failure
violated; lock it.
The invocation-limiter message uses shell-style `KEY=N` syntax
(not `key="value"`), did not exhibit the same behavior in
testing, and is left untouched.
Tested: bun test src/tools/AgentTool/specialistRouter.test.ts
17 pass (13 existing + 4 new), 61 expect() calls
Confidence: high
Scope-risk: narrow
Directive: any future change to redirect/error text returned
to the model should preserve the no-key="value" property.
* feat(orchestration): require copying project tool rules into sub-agent prompts
Live test on the desktop `+` orchestration mode showed mimo-v2.5-pro fan
out three sub-agents in parallel for a multi-task PR review prompt — but
none of the dispatched prompts mentioned the project's mandated codegraph
MCP tool. The sub-agents fell back to plain Bash + git diff + grep, so the
project's preferred code-exploration tooling went unused.
Root cause: the orchestrator's `Propagate project conventions` rule was a
soft "for example, restate codegraph rules" bullet. The model read it as
optional. Built-in specialist system prompts (code-reviewer, security-
reviewer, debugger, etc.) explicitly enumerate Bash/grep/git as inspection
tools and never mention codegraph; CLAUDE.md is loaded but the system
prompt outweighs it. The orchestrator was the only place the gap could be
closed, and it skipped it.
This commit upgrades the rule from optional to REQUIRED in the
orchestration system prompt, names the concrete check ("scan CLAUDE.md,
AGENTS.md, .claude/rules"), and tells the orchestrator to copy matching
rules verbatim. Read-only research agents (Explore, Plan) are called out
explicitly because they run with omitClaudeMd:true and absolutely depend
on the orchestrator to forward project tool conventions.
A new ORCHESTRATION_PROPAGATE_RULES_MARKER export plus a regression test
in conversations.test.ts lock the imperative phrasing so it can't be
quietly downgraded back to a soft suggestion.
Wiring is unchanged: built-in code-reviewer/security-reviewer/commit-pr
already inherit CLAUDE.md and have MCP tools available. This is purely a
prompt-strength change, only active when the user enables the desktop
`+` orchestration toggle.
Tested: bun test src/server/__tests__/conversations.test.ts (new test
plus the two existing orchestration-prompt tests pass).
Confidence: high
Scope-risk: narrow
* fix(verification-gate): preserve reminder flag across verification reset
Caught by an orchestrator-mode code-reviewer dispatch on this branch.
When a verification subagent invocation reset the edit counter, the early
return in getVerificationGateState() hardcoded reminderAlreadyFired:false,
discarding any reminderFired flag we'd accumulated walking back from the
tail. Concretely: if a verification_gate_reminder fired AFTER the verify
(i.e. newer than verify in the transcript), the reverse walk would set
reminderFired=true on its way to the verify boundary, then throw it away
on the early return. The gate would then inject a fresh reminder every
turn while the post-verify edit count stayed above the threshold,
ballooning the prompt.
Fix: carry the accumulated reminderFired through the early return. The
reminder we saw is necessarily newer than this verify (we're walking
backwards from the tail and only got here after seeing it), so it's the
valid signal that the model was already nudged for the current edit batch.
Tests: extend verificationGate.test.ts with a "preserves a reminder fired
AFTER a verification reset (no spam loop)" case that fails on the old
hardcoded false and passes with the fix. The pre-existing "ignores a
reminder that fired BEFORE a verification reset" test still passes because
in that ordering the reminder is older than the verify, so reminderFired
stays false on the way back.
Tested: bun test src/utils/verificationGate.test.ts (11/11 pass).
Confidence: high
Scope-risk: narrow
* chat(i18n): rename `协调模式` to `编排模式` (zh / zh-TW)
Brings the simplified- and traditional-Chinese label for the desktop `+`
menu's orchestration toggle in line with the English (Orchestration mode),
Japanese (オーケストレーションモード), and Korean (오케스트레이션 모드)
labels — all five locales now use the same `orchestration` word family.
`协调` was accurate but vague; `编排` matches the term Chinese
developers already see in LangChain, CrewAI, AutoGen, and Anthropic docs,
and tells a user reading just the label what the toggle actually does
(plan → fan out → synthesize) rather than a generic `coordinate`.
UI-text only. No behavior, store, WS message, or i18n-key changes.
* feat(desktop): welcome-screen task cards with one-click orchestration
Adds four quick-start task cards to the new-session welcome screens — both
EmptySession (the no-session-yet entry) and ActiveSession's empty state
(when a fresh session tab has no messages yet). Each card pre-fills the
composer with a starter prompt for a common workflow:
- Pre-merge code review (auto-enables Orchestration mode)
- Investigate a failing test (auto-enables Orchestration mode)
- Write unit tests (no orchestration)
- Explain unfamiliar code (no orchestration)
Cards flagged `orchestrate: true` automatically turn on Orchestration
mode for the session that gets created (or for the live session in
ActiveSession's empty state), so a new user sees fan-out behavior on
first contact instead of having to discover the `+` menu first. This is
the intended discovery surface for the orchestration directive shipped on
this branch.
Plumbing:
- WelcomeTaskCards.tsx is a shared component used by both welcome
surfaces, so card list, copy, and styles stay in one place.
- EmptySession owns its own composer textarea, so card click flips
component state directly (setInput + draftOrchestrate flag) and the
flag is applied via setSessionCoordinatorMode after createSession()
resolves and before connectToSession() — chatStore replays the
persisted toggle on connect, so the CLI launches with the right
--append-system-prompt.
- ActiveSession's session is already live, so a card click dispatches a
`cc-haha:composer-prefill` window event that ChatInput listens for
and applies to its composer state when the sessionId matches the
active tab. Orchestration cards call setSessionCoordinatorMode
directly, which sends `set_coordinator_mode` over WS and triggers a
CLI restart with the orchestration directive — verified live: the
server logs `[WS] Restarted CLI for <id> with runtime override`
immediately after the click.
- Cards never disable an existing Orchestration toggle. Non-orchestration
cards simply leave the toggle alone, so users who pre-enabled
Orchestration via the `+` menu don't have it silently turned off.
- Hidden on phone-sized H5 viewports (composer is already dense there).
i18n: ten new keys under `empty.tasks.*` with localizations for en,
zh, zh-TW, jp, kr (the same five locales the orchestration label
already covers). The orchestration hint reads `将通过编排模式分派给
专家子代理处理` in zh — aligning with the
`协调模式`→`编排模式` rename shipped in the previous commit.
Tests: five new tests in EmptySession.test.tsx covering: cards render on
desktop, hidden on mobile, click pre-fills the composer, orchestration
cards persist coordinator mode after submit, non-orchestration cards
leave coordinator mode untouched. `bun run vitest` 22/22 pass.
`bun run lint` (tsc --noEmit) clean.
Tested live in browser: clicked the Pre-merge review card → composer
filled with the 78-char zh prompt → localStorage carries
`coordinator-modes[<sessionId>]: true` → server log shows the CLI
restarted with `--append-system-prompt`. Screenshot:
artifacts/desktop-welcome-task-cards.png.
Confidence: high
Scope-risk: narrow
* chore(scripts): one-click MCP smoke setup for Windows
Wraps the recurring 4-step setup from docs/desktop/10-local-mcp-testing.md
into a single idempotent script. Same workflow we'd otherwise type by hand
every time we want to drive the desktop UI from Chrome DevTools MCP.
scripts/dev-mcp-test.ps1 does:
1. Health-check the API server on :3456 — fail fast with the start
command if it's down.
2. Health-check Vite on :1420 — fail fast with the start command if
it's down.
3. Confirm Vite's /health proxy is wired through to the server (catches
stale vite.config.ts).
4. Ensure http://localhost:1420 is in H5 allowedOrigins (PUT only when
missing).
5. Regenerate a fresh H5 token via /api/h5-access/regenerate.
6. Build the full URL with serverUrl + forceH5=1 + h5Token, copy it to
the clipboard, and print it.
7. With -Open, also Start-Process the URL in the default browser.
8. With -Quiet, only print the URL on stdout (for piping into other
tools, e.g. un run mcp:test:open chaining).
Surfaced via two npm scripts:
bun run mcp:test # generate URL + clipboard
bun run mcp:test:open # also open in default browser
Deliberate non-goals:
- Does NOT start server / Vite. Those are persistent dev processes the
developer wants to control lifecycle for; auto-starting them from a
one-shot script makes "is it still running?" questions ambiguous and
leaves zombies on script exit. The script tells the user the exact
command to run if either is down.
- Does NOT touch Electron. This workflow is the documented browser path
for Chrome DevTools MCP smoke; the Electron path doesn't need any of
this and runs through bun run electron:dev.
- Does NOT modify vite.config.ts proxy or desktopRuntime.ts forceH5
bypass. Both are already on this branch, and validation of those two
being correctly wired is what step 3 covers.
Doc: docs/desktop/10-local-mcp-testing.md grew a new "step 0 一键脚本"
section pointing at the script and naming the two npm targets, ahead of
the original 4 manual steps so future maintainers find the script first.
Verified by running it: server + vite both up, allowedOrigins pre-existed
so no PUT needed, fresh token issued, URL copied to clipboard, browser
launched into the welcome screen with the four task cards rendering
correctly. Screenshot in artifacts/desktop-welcome-task-cards-final.png.
Confidence: high
Scope-risk: narrow
* feat(desktop): zero-token "Recent activity" panel with hand-off button
Closes the painful gap where opening a new chat in a project means
re-explaining "what was I just doing" to the model. The user explicitly
called out that they don't want the obvious fix (auto-feeding the
previous transcript or an LLM-generated summary) because that burns
tokens on every new session.
This ships a different shape: derive the activity summary on the server
from on-disk state (session JSONL + git working tree), render it as a
read-only panel on the welcome screen so the user sees it, and only
move bytes into the model when the user explicitly clicks "Continue
from here" — and even then, only a 4-5 line hand-off paragraph (~60
tokens), not the previous transcript.
Server (zero LLM, all derivation):
- GET /api/projects/recent-activity?workDir=<abs>&excludeSessionId=<id>
- projectActivityService.getRecentActivity reads:
- sessionService.listSessions to find the most recent meaningful
session (skips empty messageCount=0 sessions and any
excludeSessionId, e.g. the just-created Untitled tab in
ActiveSession's empty welcome state, so the panel surfaces the
ACTUAL previous work session).
- Streams that session's JSONL and derives:
* lastUserMessageExcerpt (160 chars max, mid-word ellipsis)
* filesEdited list — Edit/Write/NotebookEdit tool_use file_paths,
de-duplicated in first-appearance order, capped at 8.
Hard-capped at 50k JSONL lines so giant transcripts don't hang
the welcome render.
- getRepositoryContext (existing, cached) for branch + default
branch + dirty state, plus two cheap git rev-list/status calls
for ahead/behind counts and exact dirty file count.
- Strict timeouts on each git command so the panel never blocks UI.
- Wired into router.ts under case 'projects'.
Frontend:
- RecentActivityCard reads via projectsApi.recentActivity.
- Two paths into it:
1. EmptySession (sidebar "New session" → user picks workDir):
shows once a workDir is chosen. "Open this session" switches
to the previous tab; "Continue from here" prefills the
hand-off into the empty composer's textarea.
2. ActiveSession's empty welcome state (sidebar "在 X 中新建会话"
creates a session immediately): excludeSessionId={activeTabId}
skips the brand-new empty session, hideContinueSessionButton
hides "Open this session" since the user is already on an
empty tab. "Continue from here" dispatches the existing
cc-haha:composer-prefill window event so ChatInput picks it
up — same plumbing the welcome task cards use.
- Auto-refreshes every 60s — git state changes between renders
(commits in another terminal, etc.) get picked up without UI work.
- Hidden on phone-sized H5 (composer is dense enough already).
Hand-off paragraph contents (i18n'd, ~60-100 tokens):
Last session on <branch>: "<title>".
Files touched: <up-to-5-files> (+N more).
Local is ahead of upstream by N commit(s).
N file(s) have uncommitted changes.
Please continue from there. Pick up by ...(describe the next step).
Token cost summary:
- Render: 0 tokens. Pure on-disk derivation, never sent to a model.
- "Open this session": 0 tokens. Just a tab switch.
- "Continue from here": ~60-100 tokens, ONLY if the user types more
and presses send. The user sees the prefill in the textarea before
sending and can edit or delete it.
i18n: 18 new keys (empty.recentActivity.*) localized for the same five
locales as the orchestration toggle (en/zh/zh-TW/jp/kr).
Live-tested via Chrome DevTools MCP on this branch:
- Opened cc-haha welcome path.
- Server returns correct shape: branch=feat/agent-routing-hardening,
aheadCount=6, dirtyCount=13, lastSession with 10 messages and
correct title.
- excludeSessionId correctly skips the empty just-created session and
surfaces the previous "请审查我当前分支..." session.
- "Continue from here" click prefills the textarea with the 152-char
Chinese hand-off paragraph; localStorage and WS state untouched
(zero token cost confirmed).
- Screenshots: artifacts/desktop-recent-activity-card.png and
artifacts/desktop-recent-activity-handoff.png.
bun run lint passes (tsc --noEmit clean).
Confidence: high
Scope-risk: moderate (new endpoint, but read-only and isolated; new UI
panel, but additive only — empty workdir or no prior sessions just
hides the card)
Tested: live MCP smoke per above.
Not-tested: server-side unit tests for projectActivityService (planned
follow-up; live integration validated the happy path end-to-end).
* fix(desktop): compact recent-activity card so composer stays in view
Live MCP smoke caught a layout regression. Welcome screen vertical
budget on a 923px viewport is roughly:
79 tab bar
+ ~250 hero block (logo + title + subtitle)
+ ~165 task cards
+ ~222 composer (ChatInput in hero variant)
+ ~64 flex padding
≈ 780px before activity panel
The first cut of RecentActivityCard rendered a 235px column block
(title row, chips row, italic excerpt row, file-chips row, action row).
Together with the task cards that pushed total content to ~410px
beyond what fits, and ActiveSession's empty-state wrapper used
flex flex-1 + justify-center without overflow handling — so the
sibling ChatInput got pushed below viewport bottom (composer.bottom
1112 on a 923 viewport, visible:false confirmed via DOM measurement).
Two changes:
1. Compact the card to ~127px (down 108px, roughly half) by collapsing
to a single horizontal row: icon | title + chips column | action
buttons. The italic last-user-message excerpt and the file-name
chip strip both go away — they're nice-to-haves, not load-bearing.
The data they carried is preserved: filesEditedCount stays as a
chip with a tooltip showing the first 5 file basenames; the lead
excerpt is still in the hand-off paragraph the "Continue from here"
button injects, so the user gets it where it counts (in the
composer) instead of as ambient ornament on the welcome screen.
Action buttons collapse their labels under sm: breakpoints to keep
the card single-row even in narrow side panels.
2. Add min-h-0 + overflow-y-auto to ActiveSession's isEmpty wrapper.
Belt-and-suspenders: future taller content (extra cards, longer
localized strings) now scrolls inside the welcome region instead of
shoving the composer offscreen. Composer remains a flex sibling and
stays anchored at the bottom regardless of content height above it.
Verified live in MCP browser:
recentActivity height: 235 → 127px
composer.bottom: 1112 (offscreen) → 907 (visible, on a 923 viewport)
textareaVisible: false → true
Tested across two projects with very different activity profiles
(cc-haha: short title + 0 files edited; layout-editor: long title +
20 files edited + 1876 messages). Both render in single horizontal
row.
Screenshot: artifacts/desktop-recent-activity-compact.png
bun run lint passes (tsc --noEmit clean).
Confidence: high
Scope-risk: narrow (CSS / layout only, no behavioral change).
---------
Co-authored-by: 你的姓名 <you@example.com>