Pimp My IDE / continuity shop
Back to garage
October 3, 2026 | agent runtime / recovery

Resume is a restore, not a scroll.

A transcript can look intact while the workspace, last response, queued input, model, or permission mode has changed. Check the seam before the next tool call.

Restore five things together: accepted input, workspace revision, completed effects, runtime settings, and current authority.
Session recorder / local demoRestore seam found
InputStateEffectsAuthority
Checkpoint 01842 fields unchecked

The transcript is one part of the state.

Agent tools now preserve longer sessions across windows, restarts, reconnects, and compaction. That makes continuity useful. It also gives a restored transcript more authority than it deserves.

Claude Code 2.1.288 fixed several distinct resume faults. Its notes mention files and context dropped after compaction, a last response that was not saved, a transcript cut short during a concurrent rewrite, and earlier thinking lost across versions.[1] One button labeled Resume covered several storage and reconstruction paths.

A session is restored only when its conversation and its outside effects agree.

Uncertain delivery needs a third state.

OpenAI Codex 0.160.0 fixed queued messages after reconnect. The client now waits for uncertain submissions to resolve before it resumes unsent input, which avoids duplicate sends.[2] That is the right model for every action near a broken connection.

Do not force a disconnected tool call into only failed or completed. Keep an UNKNOWN state until you read the remote result, inspect the target, or prove that the action is safe to repeat. Retrying an unknown read wastes time. Retrying an unknown write can duplicate work.

Compaction changes the evidence path.

Zed 1.22 added automatic subagent compaction and model selection when a new subagent starts.[3] Compaction can keep work moving. It also means the live context is no longer the full transcript. A useful restore record names the checkpoint, the compacted summary, and the source records that remain available.

Model and runtime settings belong in that record too. Codex 0.160.0 says its terminal interface now preserves the server provider, reasoning summary, and verbosity settings across resume and fork history.[2] A restored task on a different provider or policy is a new run, even if the transcript looks familiar.

One session can span more than one screen.

VS Code 1.140 runs its Copilot harness in a dedicated agent-host process and lets several editor windows connect to the same session.[4] The release also lets users archive one finished chat inside a larger session and later restore its title and transcript.[4] Those are useful controls. They make single-writer authority and current workspace identity visible requirements.

Before a resumed session writes, show which client holds action authority. Show the exact repository, worktree, and revision. If two clients can send input, mark queued and accepted messages with IDs so the operator can tell what the runtime received.

Run a five-line restore check.

  1. Record the last accepted user input and its message ID.
  2. Record the last completed assistant turn and checkpoint ID.
  3. Pin the repository, worktree, revision, and dirty-file summary.
  4. List tool calls near the break as completed, failed, or unknown.
  5. Read back the provider, model, permission mode, and client that owns the next action.

If any line is unknown, resume in review mode. Inspect before the runtime gets another write.

Interactive makeover / resume checksum bench

Inspect the seam before re-engaging.

Traditional purpose replaced: a resume button beside old chat. Better version: choose the interruption, trace four restore checks, and copy a packet that keeps unknown effects visible.

Interruption selector

These are sample incidents. Selection changes only this page.

Observed break
Generated restore packet

Keep unknown work unknown

Selected checks define the packet structure. They do not inspect a real session.

Restore routeNETWORK DROP / 0 OF 4
Restore checks
Restore check is open.No restore field is selected. Keep writes blocked.
Four selected checks mean the packet structure is ready. They do not prove that any message, file, effect, model, permission, or controller was restored.

Sources read

Source log and evidence boundary
  1. Claude Code v2.1.288 release notes, published October 2, 2026 and read October 3, 2026. They support the listed fixes for resume, compaction, transcript writes, model restoration, duplicate tool calls, and timeout recovery.
  2. OpenAI Codex 0.160.0 release notes, published October 1, 2026 and read October 3, 2026. They support the reconnect handling for uncertain queued messages and the preserved provider, reasoning-summary, and verbosity settings.
  3. Zed 1.22.0 stable release notes, published September 30, 2026 and read October 3, 2026. They support automatic subagent compaction and model selection for newly created subagents.
  4. Visual Studio Code 1.140 release notes, published September 30, 2026 and read October 3, 2026. They support multi-window connections to a dedicated agent host and per-chat archive and restore behavior inside a session.

Evidence boundary. The checksum bench uses synthetic incidents and browser-local state. It does not connect to Claude Code, Codex, Zed, VS Code, a repository, or an agent runtime. The five-line restore check is Pimp My IDE's editorial proposal.