Fix compression-exhausted stream finalization

[franksong2702]

Thinking Path

• Long single-turn, tool-heavy sessions can exhaust context after Hermes Agent fails to compress effectively.
• WebUI must distinguish streamed interim/progress text from a real final assistant answer.
• A persisted transcript ending in a tool result, assistant tool-call turn, or internal context-compaction reference marker is not a completed answer.
• The UI should surface compression exhaustion as an error and keep internal reference-only summaries out of the settled transcript.

What Changed

• Classifies compression exhaustion errors from agent/provider result text.
• Treats failed, partial, compression_exhausted, tool-tail transcripts, assistant tool-call tails, and context-compaction marker tails as terminal failures instead of completed turns.
• Removes the _assistant_added short-circuit so final-answer validation always checks the persisted transcript.
• Keeps [CONTEXT COMPACTION — REFERENCE ONLY] content out of settled transcript rendering while preserving transient running compression status.
• Adds regression coverage for terminal failure detection, context-compaction marker filtering, and final-answer semantics.
• Updates the changelog for the user-visible behavior change.

Why It Matters

This prevents long tool-heavy sessions from appearing completed when Hermes Agent stopped before writing a final assistant answer. It also prevents internal context-compaction reference text from being rendered as user-facing final content.

Related to #3315, NousResearch/hermes-agent#36624, and NousResearch/hermes-agent#36626.

Verification

• python -m pytest tests/test_auto_compression_terminal_failure.py tests/test_auto_compression_card.py tests/test_issues_373_374_375.py tests/test_issue765_streaming_persistence.py::TestIssue765FollowupHardening::test_silent_failure_path_does_not_reacquire_agent_lock -q
• node --check static/ui.js
• node --check static/messages.js
• git diff --check

Risks / Follow-ups

• This is the WebUI companion to the Hermes Agent compression fix; it does not by itself reduce agent-side context size.
• The change intentionally avoids rendering internal compaction reference text in settled transcripts, while still allowing transient compression status during active runs.

Contract Routing

• Contract family: runtime streaming finalization and session transcript visibility.
• Evidence: focused regression tests for terminal failure detection, final-answer semantics, and settled transcript rendering.
• Contract change: none intended; this restores the invariant that completed UI state requires a real final assistant answer.

Model Used

AI-assisted implementation with OpenAI GPT-5 Codex in a local coding workflow. The assistant inspected repository code, wrote targeted tests, implemented the fix, and ran the verification commands above.

[nesquena-hermes]

Triage: hold + changes-requested — thanks @franksong2702. The goal is right (surfacing compression-exhaustion as a real error instead of a falsely-"completed" turn) and the classifier itself is sound — I verified _session_lacks_final_assistant_answer() empirically: it correctly returns HAS-final-answer for a normal completed turn and a tool→final-text turn, and terminal-failure only for tool-result/tool-call/empty-assistant/marker tails. I picked this up to advance toward release and ran it through the full gate (test suite + Opus + the Codex regression gate). Opus cleared it, but the Codex gate caught a state-consistency ordering bug on the compression path that I then confirmed against the code, so I'm holding it rather than shipping.

(CORE) Terminal-failure handling can run BEFORE the compression session-id migration, leaving frontend/backend session state inconsistent when compression exhaustion fires after the agent rotated session_id.

The new terminal-failure check is at api/streaming.py:5439-5443:

_terminal_failure = (
    _agent_result_terminal_failure(result)
    or _session_lacks_final_assistant_answer(_all_result_messages)
)
if _terminal_failure:
    _assistant_added = False
if _terminal_failure or (not _assistant_added and not _token_sent):
    ... # emits the apperror + returns

But the compression session-id migration + snapshot preservation block runs much later, at api/streaming.py:5680+:

_preserve_pre_compression_snapshot(s, old_sid)   # ~5680
... # migrate locks/cache, register continuation, emit `compressed`, save against new sid

Hermes Agent rotates agent.session_id during compression (agent/conversation_compression.py:505-520); WebUI only mirrors that rotation in the 5680+ block. So on a compression-exhausted result that arrives after the rotation, the terminal-failure path at 5443 emits the error + returns before 5680 ever runs — which means:

• _preserve_pre_compression_snapshot() is skipped → the pre-compression history may not be archived
• continuation/session migration is skipped → the error transcript is saved against the old WebUI session id
• the frontend apperror path appends a synthetic error to the old activeSid, not the migrated continuation session
  → frontend/backend session state diverges (this is the same compression-rotation subsystem where we recently held fix: keep gateway context visible in chat transcripts #3300 for a transcript-loss regression, so we're being extra careful here).

Suggested fix (needs your design call — I didn't want to hot-patch compression-rotation ordering under release pressure):

1. Factor the compression-rotation side-effect block (_preserve_pre_compression_snapshot + lock/cache migration + continuation registration) into a helper and run it BEFORE any return from the terminal-failure apperror path — OR move the terminal-failure check below the compression migration.
2. Persist the terminal error on the migrated continuation session, and in the frontend (static/messages.js apperror handler) adopt the settled/migrated session like the done path does, instead of only pushing a local synthetic error into the old active session.
3. Add a regression test for: compression exhausted AFTER session-id rotation → assert the snapshot is preserved, the continuation session is registered, and the error lands on the migrated session.

Everything else is good — the compression_exhausted classification, the label cascade, and the frontend clear-compression-UI handling are all correct. One small non-blocking note from Opus: _classify_provider_error only inspects the error string, so if a future agent path sets result['compression_exhausted']=True with an empty/non-matching error message it falls back to the generic "No response from provider" label (your included test sets both fields, so the current path is covered).

The rest of this session's transcript/streaming fixes (#3102 edit-replay, #3321 recovery-control filter) already shipped, so this isn't a wholesale rejection — just this one ordering interaction to sort out. Happy to pair on the helper-extraction if useful. No rush.

[greptile-apps]

Greptile Summary

This PR fixes streaming finalization for sessions where Hermes Agent failed before producing a final assistant answer (compression exhaustion, tool-tail transcripts). It moves compression migration logic before terminal-failure detection, adds _agent_result_terminal_failure and _session_lacks_final_assistant_answer helpers, enriches apperror payloads with session data before enqueue, and suppresses [CONTEXT COMPACTION — REFERENCE ONLY] markers from the settled transcript.

• api/streaming.py: Compression session rotation is now detected and migrated (session ID, lock, agent cache) before terminal-failure classification; silent-failure apperror payloads now carry a full session snapshot and correct old_session_id / new_session_id fields before enqueue.
• static/messages.js: The apperror handler now uses old_session_id/new_session_id fields for session routing instead of activeSid, replacing S.session and S.messages from the server payload when a session rotation occurred.
• static/ui.js: _isContextCompactionText is extracted as a standalone helper; the settled-transcript render path unconditionally skips compaction marker messages and suppresses them from reference-card selection via _shouldShowSettledCompressionReference.

Confidence Score: 4/5

Safe to merge with the minor background-error routing note addressed; core logic for compression detection, session migration, and settled-transcript filtering is correct.

The compression migration ordering, terminal-failure classification, payload enrichment ordering, and settled-transcript suppression are all logically sound and covered by focused regression tests. The one open concern is that trackBackgroundError is called with the pre-rotation session ID when the user has navigated away during a compression-exhausted failure, which could direct the background-error notification to the archived snapshot rather than the continuation session where the error was persisted. This does not affect in-view users or the persistence of the error message itself.

static/messages.js — the trackBackgroundError call in the background-stream error path

Important Files Changed

Filename	Overview
api/streaming.py	Moves compression migration before terminal-failure detection, adds _agent_result_terminal_failure and _session_lacks_final_assistant_answer helpers, enriches apperror payloads with session snapshot and session IDs before enqueue; includes new _classify_provider_error branch for compression_exhausted.
static/messages.js	Rewrites the apperror routing to use old_session_id and new_session_id fields for session matching, replacing S.session from the server payload on compression rotation. Adds compression_exhausted label and clearCompressionUi call.
static/ui.js	Extracts _isContextCompactionText helper, adds _shouldShowSettledCompressionReference guard to suppress compaction markers from the settled reference card, and unconditionally skips compaction marker messages in the main render loop.
tests/test_auto_compression_terminal_failure.py	New regression test file covering terminal failure detection, session rotation snapshot, payload enrichment ordering, apperror routing logic, and context compaction marker handling.
tests/test_auto_compression_card.py	Updated to reflect the new rendering behaviour where compaction markers are always skipped and _shouldShowSettledCompressionReference guards the reference card.
tests/test_issue765_streaming_persistence.py	Updated silent-failure guard string to match the refactored _terminal_failure combined guard.
CHANGELOG.md	Adds accurate user-visible entries for both the terminal-failure finalization fix and the suppression of compaction reference markers.

Sequence Diagram

sequenceDiagram
    participant UI as Browser
    participant Server as streaming.py
    participant Agent as Hermes Agent
    UI->>Server: "POST /stream session=A"
    Server->>Agent: run_conversation()
    Agent-->>UI: stream delta tokens
    Agent-->>Server: "result with failed=True and compression_exhausted=True"
    Server->>Server: Detect session rotation A to C
    Server->>Server: Preserve snapshot and migrate session object
    Server->>Server: _agent_result_terminal_failure returns True
    Server->>Server: Persist error message on session C
    Server->>Server: "Enrich apperror payload with old_session_id=A"
    Server-->>UI: "SSE apperror type=compression_exhausted old_session_id=A new_session_id=C"
    alt User still on session A
        UI->>UI: "eventSid=A matches currentSid"
        UI->>UI: Replace S.session with continuation session C
        UI->>UI: Update URL and render error bubble
    else User navigated away
        UI->>UI: eventMatchesCurrent is false
        UI->>UI: trackBackgroundError for session A
    end

Loading

_{Reviews (6): Last reviewed commit: "Merge origin/master for PR #3316 conflic..." | Re-trigger Greptile}

[franksong2702]

Pushed follow-up 473fb51a plus merge updates to current master.

What changed:

• Addressed the P1 static/messages.js review: the apperror session match no longer treats activeSid as a match fallback. It now applies an error to the visible session only when the payload session id or continuation session id matches the currently loaded session, so background stream failures route through trackBackgroundError instead of attaching to an unrelated session.
• Addressed the payload-ordering review in api/streaming.py: terminal apperror payloads are now enriched with the session snapshot and session ids before put('apperror', ...), so enqueue no longer relies on mutating the same dict later.
• Added one extra companion guard while reviewing the worker patch: the generic exception apperror path now includes session_id / old_session_id before enqueue as well. This keeps the stricter frontend matching compatible with existing non-compression errors.
• Left _agent_result_terminal_failure() partial semantics unchanged. For this PR, partial remains a non-done terminal result unless we separately verify a different Hermes Agent contract.

Verification:

• node --check static/messages.js
• ./.venv/bin/pytest tests/test_auto_compression_terminal_failure.py tests/test_auto_compression_card.py tests/test_issue765_streaming_persistence.py tests/test_issues_373_374_375.py -q -> 111 passed
• git diff --check

PR is now mergeable again after merging current origin/master; GitHub checks are running on head a5f13e1e.

AI assistance: implemented by Codex sub-agent gpt-5.3-codex-spark, reviewed and adjusted by Codex GPT-5.

[franksong2702]

Addressed the new Greptile P1 on the exception path in 8ad4dc3d.

What changed:

• In the generic exception apperror path, session_id remains the persisted/current session id via getattr(s, 'session_id', session_id).
• old_session_id now uses the original _run_agent_streaming(..., session_id=...) argument, so users still viewing the pre-compression session can match the inline error after an agent-side session rotation.
• Strengthened the regression test so it asserts the exact routing assignment, not just that both fields exist.

Verification:

• ./.venv/bin/pytest tests/test_auto_compression_terminal_failure.py tests/test_auto_compression_card.py tests/test_issue765_streaming_persistence.py tests/test_issues_373_374_375.py -q -> 111 passed
• git diff --check

AI assistance: Codex GPT-5 reviewed the bot comment, made the targeted fix, and re-ran the focused regression gate.

[franksong2702]

Updated with merge commit e50e6b0 to bring the branch onto latest master and resolve the CHANGELOG conflict while preserving the compression-exhausted finalization fixes. Local verification: git diff --check; pytest tests/test_auto_compression_terminal_failure.py tests/test_run_journal_frontend_static.py tests/test_webui_state_db_context_reconciliation.py tests/test_issue2481_selected_text_reply.py tests/test_readme_compat_section.py -q (25 passed).

[franksong2702]

Conflict cleanup pushed in c2c7c04b.

What changed:

• Merged latest origin/master into franksong2702/fix-auto-compression-tool-heavy-streams.
• Resolved the CHANGELOG.md conflict by keeping the PR's compression-exhausted finalization notes in Unreleased and preserving current release notes.
• No behavior-code conflicts were needed; the PR diff against current master remains focused on compression-exhausted terminal handling and settled compression-reference rendering.

Verification:

• python -m pytest tests/test_auto_compression_terminal_failure.py tests/test_run_journal_frontend_static.py tests/test_webui_state_db_context_reconciliation.py tests/test_issue2481_selected_text_reply.py -q -> 25 passed
• node --check static/messages.js static/ui.js
• git diff --check
• GitHub Actions: 11/11 passed on c2c7c04b

AI assistance: Codex coordinated the conflict cleanup, delegated the merge cleanup to a sub-agent, then reviewed the diff and reran focused verification before pushing.

[franksong2702]

Conflict cleanup pushed in 08c501d0.

What changed:

• Merged latest origin/master (4c545a33) into franksong2702/fix-auto-compression-tool-heavy-streams.
• Resolved the CHANGELOG.md conflict by keeping this PR's compression-exhausted / tool-tail finalization notes in Unreleased and preserving current release history.
• No manual runtime/static conflict resolution was needed; the diff against current master remains scoped to compression-exhausted terminal handling and settled compression-reference rendering.

Verification:

• git diff --check origin/master..HEAD -> passed
• node --check static/messages.js static/ui.js -> passed
• /Users/xuefusong/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_auto_compression_terminal_failure.py tests/test_auto_compression_card.py tests/test_issue765_streaming_persistence.py tests/test_run_journal_frontend_static.py tests/test_webui_state_db_context_reconciliation.py tests/test_issue2481_selected_text_reply.py tests/test_readme_compat_section.py -q -> 93 passed
• GitHub Actions on 08c501d0 -> 11/11 passed
• GitHub mergeability -> MERGEABLE / CLEAN

AI assistance: Codex coordinated the conflict cleanup with a sub-agent, then re-reviewed the streaming/static diff and reran focused verification before pushing.

[franksong2702]

Marking this PR as Draft to align with the current hold / changes-requested state and reduce reviewer confusion while #3401 is still active.

This remains a valid live-to-final terminal/no-final corner case: compression-exhausted or tool-heavy runs must not settle as a normal completed answer when no real final assistant answer exists. However, #3401 now overlaps some of the same streaming finalization guards, so keeping this PR ready before #3401 lands risks reviewing the same lifecycle logic in two places.

Plan:

• Keep Fix compression-exhausted stream finalization #3316 open as the compression-exhausted / no-final follow-up.
• Wait for Redesign live-to-final assistant replies #3401 to land or otherwise settle.
• Rebase this PR afterward and cut it down to only the remaining compression-specific invariants, especially session-rotation/error-routing coverage that Redesign live-to-final assistant replies #3401 does not already own.

Refs #3401 and #3315.

[nesquena-hermes]

Shipped in v0.51.292 🎉 Thanks @franksong2702 — you fixed the compression-rotation ordering bug I held on (the migration + pre-compression snapshot + continuation registration now run BEFORE the terminal-failure return, so state stays consistent when exhaustion fires after session_id rotation) and refined the classifier so a streamed answer (_token_sent) is never misclassified as terminal-failure. Re-gated fresh on a 3-way merge onto current master: full suite 7993, Codex SAFE + Opus SAFE (both verified the migration-before-failure ordering, no false-positive on legit completions, error persists on the migrated continuation session, and [CONTEXT COMPACTION] markers are filtered from settled cards). Closes #3315.