fix(desktop): preserve full transcripts for running sessions

[Kinkoolino-Hermes]

What does this PR do?

Keeps the complete persisted Desktop transcript visible when a warm session is reopened during a running, context-compressed turn.

session.activate.messages is the agent's runtime context projection, so it may intentionally omit older durable conversation rows. The warm-cache path previously applied the fuller REST transcript only while idle. During running: true, reopening the session therefore replaced the visible conversation with the compressed runtime projection.

This change always treats a matching REST transcript as the display authority, then projects the live inflight/queued tail onto it. A three-way, id-based overlay also preserves stream deltas and local submissions that arrive while activation is waiting for REST.

The dedupe is deliberately boundary- and structure-aware rather than global text matching: REST suppresses inflight.user only when the matching user row is the open persisted tail, while local rows are considered activation-owned only when they form the expected synthetic user/assistant-stream pair. Ambiguous rows are retained so an accepted prompt is never dropped. Intentionally repeated prompts remain distinct.

Related Issue

Fixes #67871

Related, but not duplicate:

• fix: preserve queued prompt boundaries end to end #63298 addresses queued prompt identity end to end; it does not restore older persisted rows in this warm activation path.
• [Bug]: Desktop stays stuck on stale "Thinking" state after turn completes and transcript is persisted #50159 covers a stale completed-turn/Thinking view; this PR covers compressed-history replacement and duplicate in-flight user rows while running: true.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• apps/desktop/src/app/session/hooks/use-session-actions/index.ts
  • apply the matching persisted transcript while a warm session is running, not only while idle;
  • keep the matching REST transcript as the display authority on cold session.resume with a compressed live projection;
  • reconcile the live activation projection back onto that complete transcript;
  • overlay cache rows changed during the REST wait by stable local message id;
  • overlay runtime-cache rows changed before a cold resume becomes the active view;
  • avoid converting the full persisted transcript more than once in this path.
• apps/desktop/src/app/session/hooks/use-session-actions/utils.ts
  • add conservative, boundary-aware inflight.user deduplication;
  • preserve the live prompt when histories have no safe common anchor;
  • remove only structurally matched synthetic activation rows;
  • overlay only cache rows that changed relative to the pre-activation baseline.
  • replace the activation assistant placeholder with the real time-based stream id instead of rendering two pending assistant bubbles.
• Regression tests cover:
  • compressed runtime history versus a fuller persisted transcript while running: true;
  • partial assistant output remains visible;
  • an already-persisted in-flight user row appearing exactly once;
  • intentionally repeated prompts remaining separate;
  • a compressed cache's represented optimistic row not being replayed;
  • unmatched and text-identical local race prompts remaining intact;
  • assistant deltas and local submits arriving during a delayed REST request;
  • an answered historical repeat not suppressing a new identical in-flight prompt;
  • unchanged cache rows never overriding a newer authoritative REST row.
  • a full renderer restart retaining older persisted rows alongside inflight and queued.
  • a real stream id differing from the activation projection still producing exactly one pending assistant bubble.
  • a cold-resume runtime delta arriving while REST is pending remaining in the final stream bubble.

How to Test

1. Run the focused regression suite:

   cd apps/desktop
   npx vitest run --project ui \
     src/app/session/hooks/use-session-actions.test.tsx \
     src/app/session/hooks/use-session-actions/utils.test.ts

2. Run the full Desktop gate:

   npm run check

3. Manual behavior:
   • open a long session whose runtime history has been compacted;
   • start a new turn, switch away, then switch back while it is running;
   • verify older persisted messages remain visible, partial assistant output remains live, and the current user row appears once.

Verified locally

• Focused UI tests: 66 passed
• Desktop TypeScript typecheck
• Scoped ESLint and Prettier checks
• Complete final-head Desktop gate (npm run check): exit 0, including the full Vitest matrix, Desktop platform/package matrix, and production build

Safety / Regression Notes

• No session data, persistence schema, gateway payload, configuration, or backend behavior changes.
• The warm cache is still painted before network reconciliation, so this does not delay the first visible session switch.
• REST failure still falls back to the existing runtime/cache projection.
• Persisted data is accepted only when its session id matches the activated stored session, preserving the existing cross-wiring guard.
• The same persisted-authority rule applies to warm activation and cold resume; watch-window resumes without REST prefetch keep their existing path.
• Concurrent cache updates are reconciled by local message id against the captured baseline; unchanged stale cache rows cannot replace newer REST rows.
• A changed real assistant-stream-* row replaces the one baseline-new activation placeholder; it is not appended as a second assistant bubble.
• Cold resume captures a per-runtime cache baseline when the RPC resolves and overlays later cache changes before committing the view.
• The reconciliation is linear over the converted transcript and uses no global content dedupe.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass — N/A: Desktop-only TypeScript change; the complete Desktop gate is listed above
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Linux/jsdom renderer tests, Electron platform tests, and production build

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — N/A: no user-facing API or configuration change
• I've updated cli-config.yaml.example if I added/changed config keys — N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — renderer-only behavior with Electron platform suite coverage
• I've updated tool descriptions/schemas if I changed tool behavior — N/A

Screenshots / Logs

No screenshot is needed: the bug and fix are covered by a deterministic warm-session activation test using divergent runtime and persisted histories. The original live transcript contains private conversation data and is not included.

[alt-glitch]

This was generated by AI during triage.

Related to #67871, #63298, and #50159: this Desktop display-layer repair restores persisted history during a running compressed session and deduplicates the live tail; the linked work addresses the report, queued-prompt identity, and stale completed-turn mechanisms respectively.

[Kinkoolino-Hermes]

Refreshed this branch onto current main (07e97d2f5) and resolved the Desktop activation overlap deliberately rather than taking either side wholesale:

• preserved upstream's deferred REST-prefetch paint and crash-survivable in-flight journal recovery;
• preserved this PR's full persisted-transcript authority plus live inflight/queued projection reconciliation;
• preserved runtime-cache changes that arrive while cold resume is waiting;
• retained the pre-recovery transcript fail-latch.

Verification on the refreshed head:

• focused session-action suites: 87 passed;
• full Desktop UI gate: 2,210 passed, 1 skipped;
• Electron/platform gate: 725 passed, 2 skipped;
• full Desktop build and unpacked packaging: exit 0.

New head: 1577328409205f8256c00ae719a52d31d4d4c625.

[teknium1]

Thanks for covering both warm activation and cold resume; the current-main premise is real: apps/desktop/src/app/session/hooks/use-session-actions/index.ts:742-747 only applies the persisted transcript while idle, and :921-934 bypasses it when a live projection exists.

Problems

• apps/desktop/src/app/session/hooks/use-session-actions/utils.ts:479-484 clears inflight.user after finding the persisted row. That also removes the signal appendLiveSessionProjection uses at current main utils.ts:485-494 to insert an assistant boundary before queued.user. With a persisted in-flight user, queued prompt, empty assistant text, and streaming: false, this produces adjacent user rows rather than preserving the two-turn boundary. The new tests cover the queued case only with an assistant payload and streaming: true.

Suggested changes

• Keep the source in-flight-user signal while suppressing only its duplicate bubble, and add the no-delta queued-turn regression.
• Integrate manually over current main's post-base reconciliation work, notably 2109a1875's redirect-correction support.

Automated hermes-sweeper review.

[teknium1]

Clearing inflight.user also removes the (inflightUser && queuedUser) signal that appendLiveSessionProjection uses to retain an empty assistant boundary before a queued prompt. Keep dedupe state separate from the source user, and cover the persisted-user + queued-user + no-assistant-delta case.

[Kinkoolino-Hermes]

Addressed the queued no-delta turn-boundary finding and refreshed the branch onto current main (8defb9fd6).

• Kept the source inflight.user signal after persisted-tail validation while suppressing only its duplicate rendered bubble.
• This preserves the empty assistant boundary before queued.user when the assistant payload is empty and streaming: false.
• Added the missing regression through the real dedupe → projection path.
• Retained current-main redirect-correction reconciliation from 2109a1875.

Local verification on head d1cd0c1fc9fcb08cf4f00d88d4b92105e35b768b:

• regression test: RED before the fix, GREEN after it
• focused session-action utilities: 52 passed
• Desktop UI: 3,139 passed
• Electron: 867 passed, 2 skipped
• Desktop typecheck, focused ESLint, Prettier, and git diff --check: passed
• full Desktop build/package gate (test:desktop:all): passed

[GottZ]

This was generated by AI during triage.

Summary

Four PRs are listed in this complex. #67879 directly changes the REST-versus-compressed-runtime reconciliation behind #67871, while #68176, #70720, and #71121 address the adjacent image-attachment persistence and restoration failures in #68092/#70772.

Related pull requests

• fix(desktop): preserve full transcripts for running sessions #67879 best fix — (+933/-45) — n/a: The diff makes the complete persisted transcript authoritative during running activation and cold resume, overlays concurrent live changes, and suppresses an already-persisted in-flight user without collapsing intentional repeats. It explicitly addresses teknium1’s COMMENTED keep_open review by retaining the source inflight-user signal through a private dedupe marker and adding the persisted-user + queued-user + empty-assistant regression, but contributor re-review remains appropriate.
• fix(desktop): preserve image attachments across session switches #68176 [closed] fixes — (+120/-0) — n/a: This preserves cached attachmentRefs when a matching user turn is reconciled after a same-process session switch, but it does not persist those refs for cold restoration or change compressed-transcript authority. It remains relevant as a closed predecessor explicitly superseded and incorporated by fix(desktop): keep attached images renderable across session switches and restarts #71121.
• fix(desktop): persist @image: refs instead of vision-enrichment text so attachments survive session switch and restart #70720 [closed] partial — (+369/-58) — n/a: This persists @image refs and adds attachment-aware hydration, deduplication, ordering, and rendering, but its visible diff does not handle native-vision list payloads or quote paths containing spaces. It remains relevant as a closed predecessor superseded by fix(desktop): keep attached images renderable across session switches and restarts #71121, which carries the approach forward and closes those documented gaps.
• fix(desktop): keep attached images renderable across session switches and restarts #71121 [merged] best fix — (+892/-61) — n/a: This is the merged reference implementation for Desktop: image attachments disappear after switching sessions #68092/[Bug]: Attached image renders smaller with a border after session switch / restart #70772, incorporating fix(desktop): preserve image attachments across session switches #68176 and fix(desktop): persist @image: refs instead of vision-enrichment text so attachments survive session switch and restart #70720 while adding native-vision persistence, path quoting, cold-reload E2E coverage, and corrected attachment rendering. Its diff does not implement [Bug]: Running compressed Desktop sessions replace the complete transcript #67871’s persisted-transcript authority over a compressed running projection.

Duplicates

#68176 and #70720 substantially overlap on attachment preservation and were both superseded by #71121; none is a duplicate of #67879’s compressed-running-transcript fix.

Suggested consolidation

Keep #67879 open with a salvage path: preserve its full-transcript authority, live/concurrent overlay, boundary-aware dedupe, and queued no-delta regression, then obtain contributor re-review of the change that addresses teknium1’s keep_open finding. Keep #68176 and #70720 closed as superseded by #71121; #71121 is already integrated for the separate attachment issues and should not be treated as resolving #67871.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I67871(["issue #67871 (open)"])
    P67879["PR #67879 (open)"]
    P67879 -->|best fix| I67871
    class I67871 open
    class P67879 open
    class P67879 best
    class P67879 target
    click I67871 "https://github.com/NousResearch/hermes-agent/issues/67871"
    click P67879 "https://github.com/NousResearch/hermes-agent/pull/67879"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 4 pull requests and 3 issues in this complex. Each diff was read against this issue; Assessment working set: 146 kB of PR diffs, 32 kB of issue/PR text, 12 kB of discussion (16 comments), 11 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[Kinkoolino-Hermes]

@teknium1 The queued no-delta turn-boundary finding from your review has been addressed on the current head by retaining the source inflight.user signal while suppressing only its duplicate rendered bubble. The requested persisted-user → empty-assistant → queued-user regression now passes through the real dedupe and projection path, and the current CI is green. Could you take another look when convenient?

— 🤖 Bot comment · LLM: gpt-5.6-sol · not written by a human

[teknium1]

Thank you for this fix! It was salvaged into #86595 (cherry-picked onto current main with your authorship preserved in the commit history) and is now merged. Closing this PR since the work has landed.