fix(agent): rehydrate session cost counters on agent construction (#67762)

[DavidMetcalfe]

Fixes #67762.

Problem

agent.session_estimated_cost_usd was reset to 0.0 inside init_agent (agent/agent_init.py:2049-2061) with no read from any persisted source. After a gateway restart mid-session, the live counter would silently drop to $0.00 even though session_model_usage had the real accumulated cost.

The persisted data stayed correct, so /insights would show the truth and the live counter would lie. The bug affected every gateway-mediated surface; the only surface that masked it was the agents panel, which folds children-cost via tools/delegate_tool.py:2828 and adds the parent's never-zeroed prior total.

Fix

• Adds SessionDB.get_session_cost_summary(session_id) at hermes_state.py (~line 3266). Aggregates per-model rows into a single {estimated_cost_usd, cost_status} pair. The status uses a sticky priority ladder (actual > included > unknown > latest call's status), implemented with one SUM(CASE WHEN …) query that matches the codebase's existing conditional-aggregation pattern (agent/insights.py:373-376).
• Adds agent_init.py::_rehydrate_session_cost(agent) which init_agent now calls immediately after the existing reset block. Reads from agent._session_db if available, scopes exceptions to (sqlite3.Error, AttributeError, TypeError, ValueError), logs via _ra().logger.debug() on the scoped path. Other exceptions surface.
• Adds tests/hermes_state/test_session_cost_rehydration.py with 14 regression tests covering the reader, the helper, and fail-open behavior under transient DB errors vs. bug-class errors.

Tests

• 14/14 new tests pass
• 99/99 pass across tests/hermes_state/ and tests/run_agent/test_notice_spine.py with no regressions
• npx tsc -b . in apps/desktop/ is not affected (no renderer changes in this PR)

Reproduction

1. Start a new session (hermes chat or via the desktop app).
2. Send several turns until the session has accumulated non-zero cost (e.g., $4.27).
3. Verify the live cost counter shows $4.27. Inspect agent.session_estimated_cost_usd programmatically.
4. Restart the gateway process.
5. Resume the same session.
6. Before this fix: agent.session_estimated_cost_usd shows 0.0 until the next API call adds to it.
7. After this fix: the persisted $4.27 is rehydrated into the agent's attribute on construction, so the counter shows $4.27 + post-resume.

Known caveat

agent.session_cost_status will be overwritten by the unconditional = assignment at agent/conversation_loop.py:2321 (and the equivalent at agent/codex_runtime.py:150) on the very next API call, so the rehydrated status reverts to "latest call wins" until issue #67764 (priority ladder) also lands. The cost value persists correctly across a gateway restart.

This is documented in the helper's docstring, the call site comment, the test test_init_agent_rehydration_incremental_calls, and the issue body.

Out of scope

• Phase 2 (provider-reported actual_cost_usd): not addressed here. The actual_cost_usd column exists in the schema but is never written by core code today. Once Phase 2 lands, the priority ladder here becomes the real "actual" promotion path.
• JSON SessionEntry fallback: removed during review as speculative infrastructure with no caller. Will be re-added if a real caller emerges.

Review

Cross-vendor review pass via agy (Gemini 3.1 Pro High + Gemini 3.5 Flash Medium) ran in parallel on the patch. Findings applied:

• Removed speculative _rehydration_entry JSON fallback (no callers)
• Scoped except Exception: to (sqlite3.Error, AttributeError, TypeError, ValueError) with _ra().logger.debug() log
• Replaced 3 EXISTS subqueries + a separate SELECT 1 fast-path round-trip with a single SUM(CASE WHEN …) + COUNT(*) aggregation
• Added explicit if self._conn is None: return None guard at the top of get_session_cost_summary for type safety
• Tests added for scoped-vs-bug-class exception handling

[alt-glitch]

This was generated by AI during triage.

Fix PR for #67762. Related to #67774: both touch session cost-status summary data, while this patch uniquely wires rehydration into agent construction; they need a maintainer consolidation choice rather than a duplicate closure.

[teknium1]

Thanks for isolating the construction-time reset: current main still sets agent.session_estimated_cost_usd = 0.0 in agent/agent_init.py:2531, so the underlying restart bug is real.

Problems

• The proposed summary sums all session_model_usage rows, but that table includes auxiliary-task rows (hermes_state.py:1134). Auxiliary accounting deliberately writes those rows without touching the session summary (hermes_state.py:4920-4930), while the live accumulator is updated by the main loop (agent/conversation_loop.py:2900-2917). A restart would therefore change the live counter's cost domain.
• Delegated-child cost is folded only into the parent object (tools/delegate_tool.py:2862-2880), not persisted as a parent usage row. A parent-row summary still loses that portion after restart.

Suggested changes

• Decide and encode the intended live-cost domain, then test auxiliary-task rows explicitly.
• Persist or derive the parent-plus-child rollup before rehydrating it.
• Exercise the actual init_agent() call site with a temporary SessionDB; the current new tests call the extracted helper directly.

Automated hermes-sweeper review.

[DavidMetcalfe]

Both points are correct — verified against the source.

On auxiliary-task rows: get_session_cost_summary sums ALL session_model_usage rows for the session (WHERE session_id = :sid, no task filter). The main loop writes with task="", while record_auxiliary_usage writes with explicit task names (title generation, delegate summaries, etc.). Before restart, the live counter reflects main-loop only. After restart, it would include aux tasks — the domain changes silently.

On delegated-child cost: delegate_tool.py folds child cost into parent_agent.session_estimated_cost_usd in-memory only. No update_token_counts call, no persisted row. After restart, that portion is lost.

Minimal fix for the domain mismatch: Add AND task = '' to the WHERE clause in get_session_cost_summary. This keeps the rehydrated domain matching the live counter's pre-restart behavior. Aux-task cost is already visible in /insights via the per-model aggregation — it doesn't need to be in the live counter.

For child cost persistence: This is a follow-up. The current PR fixes the restart bug for main-loop cost. Child cost persistence requires a new mechanism (either a special task='__child__' marker in session_model_usage, or a separate rollup table). That's a separate concern from the restart bug.

For test coverage: Agreed — the current tests call _rehydrate_session_cost directly. An init_agent() integration test with a temporary SessionDB would catch the domain mismatch. I'll add one.

Suggested path forward:

1. Add AND task = '' filter to get_session_cost_summary (one-line fix)
2. Add init_agent() integration test with temp SessionDB
3. Child cost persistence → separate follow-up issue/PR

[GottZ]

This was generated by AI during triage.

Summary

Seven PRs address or reference three linked issues: #67770, #67778, and #67796 rehydrate restart-time accounting; #67774, #67790, and #67804 implement sticky cost-status semantics; and #67834 only documents a proposed desktop cost display. The diffs make #67770 the recorded best fix for #67762 and #67790 the recorded best fix for #67764, while #67765 still has no gateway-to-desktop implementation in this set.

Related pull requests

• fix(agent): rehydrate session cost counters on agent construction (#67762) #67770 best fix — (+539/-1) — n/a: Rehydrates cost immediately after the construction-time reset by summing persisted main-loop rows with task = '', and adds dedicated reader, rehydration, and auxiliary-exclusion tests. The keep-open review's auxiliary-domain objection is addressed in the current diff, but delegated-child cost remains unpersisted and the tests still invoke the helper rather than the actual init_agent() path; this remains the recorded best fix for [Bug]: agent.session_estimated_cost_usd resets to $0 on gateway restart #67762.
• fix: make cost_status sticky across API calls (#67764) #67774 [closed] partial — (+81/-8) — n/a: Adds sticky status handling across SQL, runtime, and /insights, but its SQL CASE ladder can still downgrade estimated to unknown. It remains relevant as the closed earlier implementation, superseded by the complete four-level ladder and tests in fix(agent): sticky cost_status priority ladder (#67764) #67790.
• fix: rehydrate session cost from DB on gateway restart (#67762) #67778 [closed] duplicate — (+286/-4) — n/a: Rehydrates restart-time cost from the session row, but prefers positive actual_cost_usd over the accumulated estimate, catches every exception, and mixes unrelated compression and turn-lease changes into the fix. It remains relevant only as a closed, superseded rehydration alternative to fix(agent): rehydrate session cost counters on agent construction (#67762) #67770.
• fix(agent): sticky cost_status priority ladder (#67764) #67790 best fix — (+596/-16) — n/a: Implements actual > included > estimated > unknown across both SQL persistence levels, both runtime paths, /insights, and delegate folding, with dedicated behavioral tests. It is the recorded best fix for [Bug]: cost_status overwritten on every API call across SQL, in-memory, and /insights aggregations #67764's cost_status cause, while the separately reported cost_source aggregation semantics remain outside its diff.
• fix(agent): rehydrate session accumulators from SQLite on restart (#67762) #67796 duplicate — (+44/-1) — n/a: Restores cost, source, token counters, and API-call totals from the sessions row, but maps persisted input/output values into two semantically distinct counter families and adds no tests. Its keep-open review identifies the incomplete source and counter mapping; compared with fix(agent): rehydrate session cost counters on agent construction (#67762) #67770, the salvageable breadth is the correctly modeled token/API rehydration, not the current assignments.
• fix(agent): make cost_status sticky by priority (#67764) #67804 partial — (+36/-5) — n/a: Makes the two runtime assignments and /insights aggregation sticky, but still persists raw per-call status and adds no tests. Despite its keep-open COMMENTED review, fix(agent): sticky cost_status priority ladder (#67764) #67790 already contains the persistence-layer enforcement and behavioral coverage that review requests, making fix(agent): make cost_status sticky by priority (#67764) #67804 a strict subset.
• feat(desktop): show running session cost in the desktop status bar (#67765) #67834 related — (+16/-0) — n/a: The sole diff is a draft-placeholder document, so it implements neither the gateway payload nor the desktop status-bar feature requested by [Feature]: Show running session cost in the desktop status bar #67765. Its keep-open review and contributor response preserve a salvage path: resolve the prior no-user-facing-cost policy, then replace the placeholder with an accuracy-labeled gateway-to-desktop implementation and unavailable/estimated-data tests.

Duplicates

#67774 and #67804 duplicate the sticky-status work consolidated more completely in #67790; #67774 is already closed. #67778 and #67796 overlap the restart-rehydration work in #67770; #67778 is already closed, while #67796's broader counters require semantic correction before any part is salvaged.

Suggested consolidation

Keep #67770 open with a salvage path: preserve its tested main-loop cost rehydration, add an actual init_agent() restart test, and track delegated-child persistence explicitly without reopening #67778. Keep #67790 open as the recorded best fix for sticky cost_status; close #67804 as duplicate of #67790 despite its keep-open review because #67790's diff already supplies the requested SessionDB enforcement and tests, and keep #67774 closed as superseded. Close #67796 as duplicate of #67770 despite its keep-open review, or split out only correctly modeled token/API-counter rehydration after preserving the distinct counter semantics. Keep #67834 open with a salvage path only if the product-policy decision approves an accuracy-labeled cost surface; otherwise the placeholder supplies no implementation to retain.

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
    I67762(["issue #67762 (open)"])
    I67764(["issue #67764 (open)"])
    subgraph Dup67770 ["PRs duplicating each other"]
        P67770["PR #67770 (open)"]
        P67778["PR #67778 (closed)"]
        P67796["PR #67796 (open)"]
    end
    P67770 -->|best fix| I67762
    P67770 -.->|partial| I67764
    class I67762 open
    class I67764 open
    class P67770 open
    class P67778 closed
    class P67796 open
    class P67770 best
    class P67770 target
    click I67762 "https://github.com/NousResearch/hermes-agent/issues/67762"
    click I67764 "https://github.com/NousResearch/hermes-agent/issues/67764"
    click P67770 "https://github.com/NousResearch/hermes-agent/pull/67770"
    click P67778 "https://github.com/NousResearch/hermes-agent/pull/67778"
    click P67796 "https://github.com/NousResearch/hermes-agent/pull/67796"

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 7 pull requests and 3 issues in this complex. Each diff was read against this issue; Assessment working set: 93 kB of PR diffs, 64 kB of issue/PR text, 16 kB of discussion (21 comments), 11 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.