[hartmark]

Summary

This is the 4th iteration of agentic conversation tracking (v1 = #9439, v2 = #10047, v3 = hartmark/OmniRoute#4), addressing the original deferral directly:

"it introduces a large new subsystem (agentic conversation tracking, +9639 lines) that would benefit from a design RFC/discussion first to align with the project's architecture before merging piecemeal" — @diegosouzapw on #9439

What changed since that deferral:

1. Size roughly halved: +9639 lines then → +4681/-232 across 83 files now (production-only, excluding tests: +3225/-216 across 71 files).
2. The architectural concern is resolved, not worked around. The original version stored full turn content (text, tool-call args, tool results) in its own table — a second, parallel content store duplicating what call_logs already persists. This version's conversation_turn_nodes table is identity-only: just the hash-chain id, role, and a content hash for reconnect-anchor lookups. Actual turn content (text/tool-call/tool-result) is resolved on demand from the existing call-log pipeline artifact via last_correlation_id — the same full, untruncated payloads call_logs already persists, already gated by call_log_pipeline_enabled, already covered by the existing retention/cleanup lifecycle. No second copy of conversation content exists anywhere. (See open-sse/services/conversationTurnContent.ts and the migration comments in 153_agentic_conversations.sql/154_conversation_turn_nodes.sql for the full reasoning.)

What remains, honestly stated — this is not purely UI wiring:

• Two new tables (agentic_conversations: conversation-root record; conversation_turn_nodes: identity-only hash chain)
• One new service (open-sse/services/conversationTracker.ts): resolves a stable conversation id per request via fingerprint + bounded prefix-hash matching, with a strict-growth invariant so independent single-shot requests sharing identical opening content never get merged
• A thin content-resolution helper (conversationTurnContent.ts) and a couple of new dashboard API routes/pages

That's the full new backend surface — bounded, and it now follows the same storage pattern the rest of the codebase already uses (reuse the call-log artifact system) instead of inventing a new one, which is the specific thing the RFC concern was about.

Feature recap (unchanged in substance from the original description):

• /dashboard/logs: new toggleable Conversation column.
• /dashboard/logs/timeline: requests sharing a conversation id share a timeline lane, connected by an arrow, with a configurable lane-reuse window.
• Request detail panel: new Full Conversation transcript above the raw SSE event stream — Markdown rendering, per-turn timestamps, live auto-refresh building the transcript in real time from the in-flight SSE chunk buffer while a request is still streaming.
• New /dashboard/conversations page listing conversations with 2+ turns, no-forking model (an edited/duplicated mid-history turn mints its own independent conversation instead of merging), pagination, prev/next navigation.
• Continuation detection excludes the system message from the identity anchor, since real coding-agent CLIs commonly regenerate it every request with live context (timestamp, cwd, git status) — without this, that volatility alone broke every continuation check against real traffic (confirmed live on omniroute-dev: 28 consecutive requests from one real growing session each recorded as separate turn_count=1 conversations before this fix).

Independent of and no longer stacked on feat/openai-responses-store-toggle/#10121 (already merged) or feat/responses-previous-response-id-virtualization/#10262 (separately submitted) — this PR stands alone against current release/v3.8.50.

Related Issues

• Addresses the [defer] on feat(dashboard): agentic conversation tracking with live transcript view #9439; supersedes feat(dashboard): agentic conversation tracking with live transcript view (v2, rebased+cleaned) #10047 and hartmark/OmniRoute#4.

Validation

• Focused tests: agenticConversations, conversationTracker, conversationTurnContent, conversations-active-call-log-id, conversations-tree-route-seq-param, logs-detail-partial-reasoning-chunk-split, request-timeline-lane-allocation, chatcore-log-truncation, sidebar-monitoring-reorg, sidebar-visibility, check-migration-numbering — 97/97 pass
• npm run lint — clean (pre-existing baseline debt only, no new violations)
• npx tsc -p tsconfig.typecheck-core.json — clean
• check-migration-numbering.test.ts — 0 gaps, 0 collisions (15/15)
• Production-code changes include new automated tests in this PR

Tests Added Or Updated

• tests/unit/agenticConversations.test.ts, conversationTracker.test.ts, conversationTurnContent.test.ts, conversations-active-call-log-id.test.ts, conversations-tree-route-seq-param.test.ts, request-timeline-lane-allocation.test.ts, logs-detail-partial-reasoning-chunk-split.test.ts (all new)

Coverage Notes

• New backend surface (conversationTracker.ts, conversationTurnContent.ts, agenticConversations.ts) is covered by dedicated unit tests exercising the real fingerprint/hash-chain/reconnect logic, not just mocked branches.

Reviewer Notes

• Migrations 153_agentic_conversations.sql / 154_conversation_turn_nodes.sql are purely additive new tables, no backfill needed.
• Happy to split further if any specific piece (e.g. the timeline lane-reuse UI, or the tracker service itself) would review better as its own PR — flagging that option rather than assuming this exact shape is the only way to land it.

[hartmark]

CI here is failing on pre-existing base-red on release/v3.8.50, not this PR — reproduced identically on a pristine checkout of the base tip. For correlated tracking:

• mutation-test-coverage (Fast Quality Gates): untracked, fixed in fix(quality): register visionBridge-responses-9597 covering test in stryker tap.testFiles #10264
• open-sse-typecheck (stream.ts TS2300 duplicate identifier): tracked/fixed in [TS7] fix(types): remove duplicate SSE comment import #10252 and fix(ci): clear base-reds on release/v3.8.50 (round 4) #10260
• dead-code ratchet regression: tracked in 🔴 Release branch not green: release/v3.8.50 #9985 (per that issue, ratchet drift is expected mid-cycle / rebaselined at release, not a contributor concern)
• the poolside live-catalog test failure in Unit Tests fast-path: fixed in fix(ci): clear base-reds on release/v3.8.50 (round 4) #10260

[diegosouzapw]

Thanks for the thorough writeup, and for directly addressing the earlier RFC deferral rather than just resubmitting — the identity-only conversation_turn_nodes design (resolving display content from the existing call-log artifact instead of storing a second copy) is a real fix for the original architectural concern, and the comments documenting the real-traffic edge cases you hit (system-prompt regeneration, sliding context windows, cache-aware turn edits) make the reasoning easy to follow and verify.

One blocking issue found in review: 153_agentic_conversations.sql and 153_call_logs_response_id.sql both use version prefix 153, which is already taken on the current release/v3.8.50 tip by 153_radar_local_model_state.sql (merged after this branch's fork point). This isn't cosmetic — migrationRunner.ts's runtime collision guard throws on getDbInstance() once all three files are present together, which is exactly why all 4 Unit Tests fast-path shards, Vitest (fast-path), and the migration-numbering gate are currently red on this PR (confirmed directly in the job logs: Migration version collision detected: version=153 → [agentic_conversations, call_logs_response_id]). Renumbering the two files (e.g. to 155/156) should clear all of that at once.

The dead-code gate failure (411 > baseline 409) looks like separate base drift rather than something this PR caused — release/v3.8.50's baseline was already rebaselined to 415 after this branch's fork point, but that value isn't in this branch's quality-baseline.json yet. A rebase onto the current tip should pick it up automatically.

Two smaller, non-blocking items:

• resolveConversationId() is called directly with no try/catch in the hot chat-completion path (chat.ts:616), and withChatAdmission rethrows uncaught errors rather than converting them into a sanitized response. Every other side-effect call nearby in that same function is defensively wrapped — worth doing the same here so a DB hiccup in this new, best-effort tracking feature can't turn an otherwise-working chat request into a hard failure.
• getConversationTurnTree (agenticConversations.ts:230) looks like a leftover export from the tree-vs-flat-list redesign — no callers remain outside its own file/tests.

Once the migration rename lands (and ideally the resolveConversationId guard), this looks like solid, mergeable work.

[hartmark]

Addressed both non-blocking items from review:

1. resolveConversationId() try/catch (chat.ts) — wrapped the call so a DB hiccup in this best-effort tracking feature falls back to conversationId = null and logs a warning instead of turning a working chat request into a hard failure, matching the try { ... } catch {} pattern used by the other side-effect calls nearby in the same function (e.g. recordRejectedRequestUsage, getComboByName).

2. getConversationTurnTree — confirmed it: knip's project/entry config in knip.json excludes tests/** entirely, so an export used only by tests can never register as "used" no matter how many call sites exist — this wasn't a leftover, it was an intentional test-support export, but it was permanently unwinnable against the dead-code gate as written. Rather than keep a second whole-chain query path alive solely for that, I swapped its 8 test call sites (across agenticConversations.test.ts and conversationTracker.test.ts) to the paginated getConversationTurnPage(id, { limit: 500 }) — already the dashboard's canonical query — and deleted the export. All 31 tests in both files still pass with identical assertions (they only ever read .length/.map(contentHash), so the extra seq field and page wrapper are inert for them).

Also caught and fixed pre-existing docs-sync drift on this branch (42 docs/i18n/*/llm.txt mirrors were stale against root llm.txt) — unrelated to this PR's content, just blocking the pre-commit hook.

Verified: npm run typecheck:core clean, node scripts/check/check-dead-code.mjs → 415 (matches current baseline, no regression), both test files green (31/31), t11:any-budget / tracked-artifacts / docs-sync pre-commit gates all pass.

Pushed: 6e54ce1a1

[hartmark]

Rebased onto current release/v3.8.50 (head 122611b0a). Both review items addressed and verified directly in source:

• Migration collision: renumbered to 154_call_logs_response_id.sql / 155_agentic_conversations.sql / 156_conversation_turn_nodes.sql — this landed cleanly since feat(responses): virtualize previous_response_id continuation regardless of upstream support #10262 (this repo's own 154_call_logs_response_id.sql, responsesContinuationStore.ts) is now merged upstream; the rebase folded the two into one canonical implementation instead of duplicating it.
• resolveConversationId() guard: wrapped in try/catch with a log.warn fallback (chat.ts:656) — a DB hiccup in this best-effort tracking no longer risks the chat request.
• getConversationTurnTree dead export: removed, no remaining callers.

Also found and fixed two CI failures while rebasing:

• i18n-vi-completeness.test.ts: requestLogger.columns.conversation (this PR's own new column) was missing from fr.json/vi.json.
• check-docs-sync.mjs: the docs/i18n/*/llm.txt mirrors still said "117 domain-specific files" after an earlier rebase updated the migration count but not this companion number — fixed across all 42 locales.

The remaining red (chat-cooldown-aware-retry.test.ts, "expected cooldown-aware retry wait, got 60ms") is unrelated — that test file's history has no overlap with this branch's changes, and the assertion shape (a hardcoded wait-time expectation) looks like CI-runner timing flakiness rather than a real regression.

@diegosouzapw ready for another look.