feat: group session tool calls - #466

Summary

Implements Linear issue ENG-135, “Group tool calls (Synara pattern)”, for the web session transcript.

The session timeline now groups consecutive tool calls within a turn while preserving the existing event projection and storage contracts. Ordinary single calls remain ordinary rows, non-tool boundaries split groups, failures stay standalone, and a still-growing trailing group remains visibly open until the turn settles.

Root cause

The timeline renderer treated every tool-call part as an independent row. That made tool-heavy turns noisy and prevented the transcript from communicating the higher-level work phase. The existing derived-row pipeline also needed a stable identity that survives streaming replacement, retries, deduplication, and arrival-order changes without changing the user’s local disclosure state.

Implementation

• Added deterministic work-group identity based on the canonical tool-call IDs, with a persisted per-message anchor for groups already rendered during streaming.
• Derived groups from consecutive tool calls in the same turn, splitting at non-tool entries and failed calls.
• Kept the active trailing group open with a four-row visible budget and truthful remaining-count summaries.
• Preserved the existing interrupted-turn expansion behavior and kept settled groups collapsed by default.
• Added stable hidden disclosure targets and aria-controls links so keyboard and assistive-technology users can operate summaries before and after expansion.
• Preserved the existing scroll anchoring behavior; grouping does not alter the underlying timeline scroll model.
• Split group identity into its own module so production source files remain below the repository’s 500-line limit.
• Kept the change web-local: no daemon, SQLite, ACP/JSON-RPC, HTTP/SSE, UDS, native-tool, or public API changes.

Verification

Focused checks completed:

• bunx turbo run test --filter=./web --force -- --run src/components/assistant-ui/__tests__/session-timeline.logic.test.ts src/components/assistant-ui/__tests__/timeline-scroll-anchoring.test.ts src/components/assistant-ui/__tests__/session-thread.test.tsx
  • PASS: 3 files, 131 tests (including the review regression).
• bunx turbo run typecheck --filter=./web
  • PASS.
• Targeted oxfmt and oxlint for all changed source and test files
  • PASS with zero warnings/errors.
• git diff --check
  • PASS.
• Independent GPT-5.6-Sol review at high reasoning effort
  • No actionable findings after review fixes.
• Greptile P1 review finding
  • Fixed in f0d07aa2c: work-group derivation now tracks IDs already assigned within the message, prevents retained-anchor reuse, and disambiguates a colliding fallback with a stable part suffix. The regression is in the canonical timeline logic suite and passes.
• CodeRabbit review artifact
  • .codex/artifacts/coderabbit/pr-466-review.json was refreshed repeatedly; no actionable CodeRabbit findings were available because the bot reported Review limit reached. A /review request was also posted and remained rate-limited.

Focused gate and QA notes

make gate was run twice as required for focused gates. Both runs reached the web lane but were blocked by unrelated existing failures outside this diff:

1. The first run had a single failure in web/src/systems/os/hooks/__tests__/use-window-manager-stream.test.tsx (expected null, received workspace:home); the lane otherwise passed 713/714 files. Re-running the exact test through Turbo reproduced it.
2. The second run passed lint, typecheck, and codegen but timed out in the unrelated Storybook visual-contract test src/storybook/__tests__/web-storybook-visual-contract.test.ts.

Per the task instruction, make verify and make gate-full were not run.

The isolated real-user QA walk was attempted with a fresh QA lab and was torn down cleanly. It is blocked before the transcript can be exercised because the unrelated demo seed does not compile: internal/demoseed/seed.go:79:33: db.ListPresets undefined. Evidence and teardown details are recorded in docs/qa/reports/2026-08-24-eng-135-session-groups.md.

CI monitoring

The post-push CI run was monitored through completion:

• Run 32803898802, attempt 1: web/Desktop/Preflight/Build and race shard 2/2 passed; Go (format, lint) hit its fixed 10-minute golangci-lint timeout; race shard 1/2 reported 648 failures across internal/store/globaldb.
• The failed jobs were retried once, as 32803898802, attempt 2. Go (format, lint) passed on retry, while race shard 1/2 reproduced the unrelated globaldb failure with 729 failures after 2,922 seconds. All other checks passed.
• No ENG-135 source path appears in either failing job; no out-of-scope daemon or database change was made to silence the external blocker.
• CodeRabbit has no unresolved comments but cannot approve the latest commit because its review limit is exhausted. Greptile's P1 identity-collision finding was fixed and its thread is resolved.

Risks and follow-up

• The implementation is isolated to the session timeline projection and renderer; the main residual risk is integration behavior under the repository’s existing web gate failures.
• CI should provide the authoritative full-repository validation.
• The local QA scenario remains blocked-verify until the unrelated demo-seed compile failure is fixed.

Compozy Impact Audit

• Native tools: no impact. No compozy__* tool IDs, toolsets, descriptors, schemas, digests, risk flags, capability gates, diagnostics, or CLI/API fallbacks changed; only the web session projection and disclosure renderer were checked.
• Extensibility and hooks: no impact. No extensions, hooks, skills/capabilities, tools/resources, registries, bridge SDKs, MCP sidecars, or config.toml lifecycle changed; the existing event-to-row path remains the source of truth.
• Workspace data isolation: no impact. Group identity is session-message scoped in the existing local XState timeline store; no global/workspace/session/agent ownership, storage queries, SSE/event propagation, cache keys, or cross-workspace list/read paths changed.
• Official Compozy skill: no impact. No public tool IDs, CLI paths, hook events, capabilities, extension resources, or memory/network/task semantics changed.

Closes ENG-135.