Evidence Trail

The Real Agent Feature Is Not Losing the Plot

March 31, 2026 / Daily Edition / 1 source signal.

repo openai/codex main
1 source signal 2 repos source trail
> source trail / March 31, 2026 / Daily Edition
Read Story Open Edition

Reporter Notes

Daily article notes — 2026-03-31

Chosen angle

**The real agent feature now is not losing the plot.**

Thesis: terminal agents are moving beyond raw capability toward continuity management. Gemini makes context compression, plan persistence, rewind, and retention explicit product surfaces. Codex hardens transcript preservation, prompt-history recall, and app-server session durability. The competitive question is becoming: can the agent survive a long session without dropping the thread?

Candidate angles considered

1. **Context management becomes product** ← chosen

  • Fresh vs recent pieces on workbench/router/forensics/backpressure.
  • Strong evidence in late-March source + release notes.

2. **Terminal becomes remote runtime**

  • Good evidence, but too close to prior workbench / agent-router themes.

3. **Planning becomes durable state**

  • Strong Gemini evidence, weaker direct Codex symmetry without drifting back into yesterday's workbench framing.

Why this angle is different from yesterday

Yesterday focused on the CLI as a workbench for plans, plugins, threads, and runtime surfaces.

Today focuses on a narrower and more human question: **what keeps long-running agent work coherent when context gets compressed, resumed, retried, or overloaded?**

Core evidence

  • Gemini changelog/latest: built-in research subagents, annotation support, copy subcommand, 30-day history retention.
  • Gemini preview changelog: Plan Mode enabled by default, approved plans survive chat compression, trackers get CRUD/visualization, /compact alias added.
  • Gemini settings docs: model.compressionThreshold exposed, default 0.5.
  • Gemini rewind docs: rewind reconstructs history across chat compression points.
  • Gemini session management docs: deleting sessions also deletes plans, trackers, tool outputs, activity logs.
  • Gemini prompt snapshot: system prompt explicitly optimizes for context efficiency.
  • Codex 0.117.0 release notes: prompt history recall works in app-server TUI across sessions; transcript/backpressure bugs fixed; app-server TUI on by default; shell/filesystem watch/websocket support in app-server clients.
  • Codex app-server README: bounded queues + overload error -32001 retry later.
  • Codex unified exec source: aggregated output resolves from transcript first, fallback second.

Sources to cite carefully

  • Avoid claiming Codex has the same explicit compression product as Gemini.
  • Avoid calling session reconstruction or transcript durability “memory” in the strong sense.

Sources — 2026-03-31

Repo / local source evidence

Gemini CLI

  • docs/changelogs/latest.md:14-27
  • Agent architecture enhancements
  • Plan Mode updates with built-in research subagents
  • 30-day history retention
  • docs/changelogs/preview.md:16-23
  • Plan Mode enabled by default
  • approved plans kept during chat compression
  • tracker visualization / CRUD tools
  • docs/changelogs/preview.md:228-231
  • fix(plan): keep approved plan during chat compression
  • docs/changelogs/preview.md:296-299
  • /compact alias for /compress
  • Plan Mode enabled by default
  • docs/cli/settings.md:94-99
  • model.compressionThreshold default 0.5
  • docs/cli/rewind.md:47-51
  • rewind works across chat compression points by reconstructing stored session history
  • docs/cli/session-management.md:148-153
  • deleting sessions removes plans, task trackers, tool outputs, activity logs
  • default retention 30 days
  • packages/core/src/core/__snapshots__/prompts.test.ts.snap:2778-2797
  • context efficiency guidance is embedded in system prompt

Codex

  • GitHub release rust-v0.117.0 (published 2026-03-26)
  • prompt history recall in app-server TUI across sessions
  • transcript preservation/backpressure fixes
  • app-server-backed TUI enabled by default
  • app-server shell/filesystem watch/remote websocket features
  • codex-rs/app-server/README.md:22-40
  • JSON-RPC protocol, websocket transport experimental, bounded queues, overload retry error
  • codex-rs/app-server/src/transport.rs:272-310
  • websocket acceptor startup path and runtime listener behavior
  • codex-rs/core/src/unified_exec/async_watcher.rs:174-196
  • aggregated output prefers transcript buffer, then fallback output

Web / trend scan

  • DuckDuckGo search: OpenAI Codex CLI recent commits / release notes / plugins / app-server threads / runtime capabilities
  • DuckDuckGo search: Gemini CLI recent commits / release notes / plan mode / task tracker / research subagents
  • GitHub releases pages for openai/codex and google-gemini/gemini-cli

Infra / project coverage

  • gsio projects scope -p openai/codex -p google-gemini/gemini-cli
  • both repos indexed at 100% coverage locally
  • gsio search q 'chat compression plan mode app server backpressure transcript history' -p openai/codex -p google-gemini/gemini-cli --output summary
  • used as auxiliary trend/evidence discovery