feat(sessions): index workflow runs and agents (layer 2 of session intelligence)

[ScriptedAlchemy]

Layer 2 of the session-intelligence work (builds on merged #281's git spans). Indexes Claude Code workflow runs and their agents into the per-project session store, attaching them to the parent thread and — via #281's git spans — to branches/worktrees/commits.

What you can now ask

• "what workflow runs happened in this session/thread?"
• "what workflows ran on branch X / worktree Y / commit Z?" (a run inherits its parent session's git spans)
• "show run wf_… — its result, phases, and agent roster"
• "what did agent `review:bugs` in run wf_… do?" (drill to the agent, then read its messages via message_search)

What's in it

• Schema (src/sessions/workflow_index.rs): workflow_runs + workflow_agents, version-gated migration mirroring feat(sessions): git-anchored session/LCM correlation #281; idempotent upsert; runs_for_session/run_for_id/agents_for_run + the git-span join runs_for_git_scope
• Ingest (src/sessions/workflow_ingest.rs): discovers subagents/workflows/<run_id>/ dirs during session ingest, parses run meta + result + per-agent transcripts, idempotent + incremental (mtime watermark), fail-open. Verified against real ~/.claude: 11 runs / 79 agents ingested for the authoring session
• Query: tracedecay_workflows MCP tool (3 modes: thread / git-ref / run-drill) + workflow_run/workflow_agent filters on message_search; taught as rung 7 in managing-session-context
• CLI: tracedecay sessions unfinished (ported from the codex/workflow-state-tool branch)

Verification

Cherry-picked clean onto current master (resolved one skill conflict: #280 consolidated recalling-session-context into managing-session-context, so the workflow rung moved there). clippy --all-targets clean; 29 workflow tests green; full CI matrix running.

Fittingly, this makes tonight's own 11-workflow / 79-agent sprawl natively queryable instead of hand-tracked. Design: fact 38. Recovery: session `2c51d204-3565-4a10-833d-d8fbd51620c3`, workflow `wf_d8320cfc-191`.

🤖 Generated with Claude Code

[changeset-bot]

⚠️ No Changeset found

Latest commit: b92790f

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[ScriptedAlchemy]

Rebased/amended: fixed the tool-count lockstep (master gained tracedecay_grep from merged #280, so base 99→100 → 101 with workflows) and shortened the commit subject under 72 chars. Full local nextest: 3848/3849 pass. The one local failure (multi_mcp_coordination_test::two_mcps_on_same_project_coordinate_via_sync_lock) is an environmental 30s-timeout of #280's already-merged startup-catch-up sync under this machine's heavy load — not touched by this PR (layer-2 changes no coordination/sync code); expecting it green on CI's clean runner.

[ScriptedAlchemy]

The 3 CI failures are git_watch timing flakes inherited from merged #280's auto-sync (fifty_commit_rebase_needs_one_sync et al.) — real-clock-fragile tests that pass 3/3 in isolation locally (0.6s each) but flake under CI's parallel load. Layer-2 touches no git_watch/coordination/sync code (verified). Re-ran the failed jobs.

Separately worth flagging: these git_watch timing flakes now live in master (via #280) and will intermittently redden any PR's CI — worth a dedicated determinism fix (inject a clock / assert behavior not wall-time), which the earlier test-robustness workflow diagnosed but couldn't land while git_watch was still unmerged. I can take that as a follow-up to #283.