refactor(agents): remove transcript token cost enrichment

[yyovil]

Summary

• remove AgentSessionInfo.cost / CostEstimate and delete live token/cost aggregation from agent session-info paths
• keep Codex and Claude session enrichment lightweight: use persisted native metadata when available and bounded JSONL reads only for restore/session metadata discovery
• bound Aider/Cursor/Kimi summary reads and remove stale token/cost assumptions across agent tests/docs

This is a follow-up cleanup after #1992 and #1996. It keeps the Codex thread-id lookup/restore behavior needed for legacy sessions while making dashboard/session-list enrichment safe by construction.

Related: #1991, #1935.

Tests

• pnpm --filter @aoagents/ao-plugin-agent-codex test
• pnpm --filter @aoagents/ao-plugin-agent-codex typecheck
• pnpm --filter @aoagents/ao-plugin-agent-claude-code test
• pnpm --filter @aoagents/ao-plugin-agent-claude-code typecheck
• pnpm --filter @aoagents/ao-plugin-agent-aider test
• pnpm --filter @aoagents/ao-plugin-agent-aider typecheck
• pnpm --filter @aoagents/ao-plugin-agent-cursor test
• pnpm --filter @aoagents/ao-plugin-agent-cursor typecheck
• pnpm --filter @aoagents/ao-plugin-agent-kimicode test
• pnpm --filter @aoagents/ao-plugin-agent-kimicode typecheck
• pnpm --filter @aoagents/ao-plugin-agent-opencode test
• pnpm --filter @aoagents/ao-plugin-agent-opencode typecheck
• pnpm --filter @aoagents/ao-plugin-agent-grok test
• pnpm --filter @aoagents/ao-plugin-agent-grok typecheck
• pnpm --filter @aoagents/ao-core typecheck
• pnpm --filter @aoagents/ao-core exec vitest run src/__tests__/activity-events-migration.test.ts
• pnpm --filter @aoagents/ao-web test -- src/lib/__tests__/serialize.test.ts
• selected plugin builds required by web services, then pnpm --filter @aoagents/ao-web typecheck
• pnpm lint
• git diff --check

Notes: full pnpm build is blocked locally by existing packages/notifier-macos Swift build environment error (SOURCE_DATE_EPOCH empty). Full pnpm --filter @aoagents/ao-core test is blocked locally by timeouts in activity-events-migration.test.ts; the same test file passes when run directly as listed above.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9b93dfad18

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Copilot]

Pull request overview

This PR updates the Codex agent plugin to avoid expensive JSONL streaming/parsing on request paths for sessions that are already terminal (or runtime-missing) by returning a cheap AgentSessionInfo derived from persisted restore metadata (codexThreadId / codexModel). It also adds a short-lived cache so repeated getSessionInfo() / getRestoreCommand() calls within the same refresh window don’t duplicate JSONL parsing work.

Changes:

• Add a metadata-only fast path for getSessionInfo() when codexThreadId is already persisted (especially for terminal/runtime-missing sessions).
• Add a module-level cache + in-flight dedupe for streamed Codex session data to avoid repeated parsing within a TTL window.
• Extend tests to cover “no JSONL streaming for terminal sessions” and “streamed data is cached across repeated calls”.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
packages/plugins/agent-codex/src/index.ts	Adds metadata-only getSessionInfo() path and introduces cached/deduped JSONL streaming for session data.
packages/plugins/agent-codex/src/index.test.ts	Adds coverage for metadata-only fast path and the new streamed-session-data cache behavior.

[greptile-apps]

Greptile Summary

This PR fixes a Codex-session-manager OOM path by avoiding full-file JSONL streaming for sessions that already have codexThreadId/codexModel persisted to metadata, and by replacing the unbounded streamCodexSessionData with readCodexSessionMetadata which caps reads at 100 lines / 1 MB. Cost-estimation logic is removed entirely from all agent plugins as part of this change.

• Fast path for terminal sessions: buildPersistedCodexSessionInfo returns session info directly from persisted metadata without touching the filesystem, fixing the OOM trigger for long-dead sessions that were being re-streamed on every refresh cycle.
• Bounded JSONL prefix scan: readCodexSessionMetadata replaces streamCodexSessionData, reading at most 100 lines or 1 MB instead of the entire (potentially 100 MB+) file; a new maxBytes guard is added to readJsonlPrefixLines.
• Cost estimation removed: CostEstimate interface and all per-plugin token-counting / cost-aggregation code are deleted across codex, claude-code, aider, cursor, kimicode, and opencode plugins.

Confidence Score: 5/5

Safe to merge — the OOM fix is well-scoped and the bounded-read approach correctly prevents large JSONL files from being loaded into memory.

All changes are additive safeguards (byte limits, fast paths) or pure deletions of cost-estimation code with no remaining callers. The fast path for terminal sessions is covered by new tests, and the bounded JSONL scan mirrors the existing cwd-match scan approach. The two observations are minor ordering/optimization nits that do not affect correctness for real-world session files.

packages/plugins/agent-codex/src/index.ts — specifically the byte-limit ordering in readJsonlPrefixLines and the missing early-exit in readCodexSessionMetadata.

Important Files Changed

Filename	Overview
packages/plugins/agent-codex/src/index.ts	Core change: replaces unbounded JSONL streaming with a bounded prefix scan, adds buildPersistedCodexSessionInfo fast path, and removes cost estimation. Logic is correct; minor ordering note on byte-limit break vs. newline processing.
packages/plugins/agent-codex/src/index.test.ts	Test suite updated to replace stream mocks with file-handle mocks, adds coverage for terminal fast-path, live sessions with/without model metadata, and lifecycle-driven terminal detection.
packages/core/src/types.ts	Removes CostEstimate interface and its cost field from AgentSessionInfo; doc-comment updates only.
packages/plugins/agent-claude-code/src/index.ts	Removes extractCost and cost-related JSONL fields; simplifies parseJsonlFileTail to always use a file handle with bounded reads. Behavior for tail-read logic is unchanged.
packages/plugins/agent-aider/src/index.ts	Switches extractAiderSummary from readFile to a bounded 64 KB head read via open. Correct since the first user message (#### prefix) is always near the start of .aider.chat.history.md.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[getSessionInfo called] --> B{buildPersistedCodexSessionInfo}
    B --> C{codexThreadId in metadata?}
    C -- No --> D[findCodexSessionFileCached]
    C -- Yes --> E{codexModel in metadata?}
    E -- Yes --> F[Return info from metadata no I/O]
    E -- No --> G{isTerminalSession?}
    G -- Yes --> H[Return info from metadata no I/O, summary=null]
    G -- No --> D
    D --> I{Session file found?}
    I -- No --> J[return null]
    I -- Yes --> K[readCodexSessionMetadata first 100 lines / 1 MB]
    K --> L{threadId or model found?}
    L -- No --> J
    L -- Yes --> M[buildCodexSessionInfo persists codexModel to metadata]
    M --> N[Return AgentSessionInfo]
    F --> N
    H --> N

Loading

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
packages/plugins/agent-codex/src/index.ts:158-163
The byte-limit guard fires **before** the newline-processing loop, so all complete lines that happen to land in the chunk that pushes `totalBytesRead` past `maxBytes` are silently discarded. In the worst case a batch of lines that would have been within the effective budget (accumulated in `partialLine`) is lost when the loop breaks. Moving the byte-limit check to after the newline loop ensures every complete line already decoded from the last chunk is captured.

```suggestion
      partialLine += decoder.write(buffer.subarray(0, bytesRead));

      let newlineIndex = partialLine.indexOf("\n");
      while (newlineIndex !== -1 && lines.length < maxLines) {
        const line = partialLine.slice(0, newlineIndex).trim();
        if (line) lines.push(line);
        partialLine = partialLine.slice(newlineIndex + 1);
        newlineIndex = partialLine.indexOf("\n");
      }

      if (totalBytesRead > maxBytes) {
        break;
      }

      newlineIndex = partialLine.indexOf("\n");
```

### Issue 2 of 2
packages/plugins/agent-codex/src/index.ts:321-334
`readCodexSessionMetadata` continues iterating through all 100 lines even after both `threadId` and `model` have been populated. Since `session_meta` (with the thread ID) and `turn_context` (with the model) are nearly always the first two lines of a Codex JSONL file, adding an early-exit saves unnecessary JSON parsing for the remaining lines on every refresh that takes the live path.

```suggestion
        if (entry.type === "turn_context" && typeof payload.model === "string" && payload.model) {
          data.model = payload.model;
        } else if (!data.model && typeof payload.model === "string" && payload.model) {
          data.model = payload.model;
        }

        if (data.threadId && data.model) break;
      } catch {
        // Skip malformed lines
      }
    }
  } catch {
    return null;
  }

  return data.threadId || data.model ? data : null;
```

_{Reviews (3): Last reviewed commit: "refactor(agents): remove session token c..." | Re-trigger Greptile}