Summary

startOrResumeThread in the Codex app-server harness uses dynamicToolsFingerprint to decide whether to resume an existing thread or start a new one. The fingerprint is computed as JSON.stringify(sortedDeepClone(dynamicTools)) — hashing the full tool descriptor array including descriptions and runtime-populated schema enums.

Several tool specs are rebuilt per-request with current runtime context baked in (current channel name, channel-supported actions, subagent-caller flag), so the fingerprint changes between turns of the same OpenClaw session whenever the caller's surface changes. Result: Codex threads are discarded and recreated on every cross-surface interaction, destroying conversation continuity.

Also: any OpenClaw upgrade that tweaks a built-in tool description invalidates every existing Codex thread on the host.

Environment

• OpenClaw 2026.4.10 (reproduces on current installed version; behavior exists since Codex harness was introduced)
• Codex app-server via stdio transport (default)
• Issue observed across multiple machines; root cause analyzed with on-disk sidecar evidence

Reproduction (observational, on-disk)

No live reproduction needed — the bug is visible from existing sidecar files.

For any agent that has run under Codex harness across >1 OpenClaw surface or OpenClaw version, inspect:

~/.openclaw/agents/<agent-id>/sessions/*.codex-app-server.json

Each sidecar's dynamicToolsFingerprint is the JSON-stringified tool catalog snapshot at binding time. Diffing two sidecars from the same agent shows non-structural differences:

Example diff (3 sidecars from one bench agent, same machine, different runs)

session A (OpenClaw version N):    tool_count=41, hash=677be5fb...
session B (OpenClaw version N+1):  tool_count=31, hash=249f9666...
session C (OpenClaw version N+1):  tool_count=31, hash=249f9666...

Deltas observed:

1. message tool description embeds current channel name

- ...Current channel (telegram) supports: delete, edit, poll, react, send, topic-create, topic-edit.
+ ...Supports actions: send, broadcast, poll, react, delete, edit, topic-create, topic-edit.

The second form renders when no channel inbound context is present (cron, subagent, internal trigger). Same session, different triggers → different fingerprint.

2. message tool inputSchema size changes with channel: 7395 B vs 6180 B (action enum and allowed fields rebuilt per-channel).

3. Tool count delta: 41 → 31 when isSubagentSessionKey(sessionKey) flips Dropped tools when running as subagent:

gateway, sessions_spawn, cron, subagents, session_status,
sessions_send, nodes, sessions_list, agents_list, sessions_history

4. OpenClaw upgrade changed built-in tool description text

- ...Use process whenever you need logs, status, input, or intervention. Use pty=true...
+ ...Use process whenever you need logs, status, input, or intervention. Do not use exec sleep or delay loops for reminders or deferred follow-ups; use cron instead. Use pty=true...

Every upgrade that touches any built-in tool description invalidates all Codex thread bindings on the host.

Expected behavior

Same OpenClaw session should resume the same Codex thread across:

• Cross-channel calls (Telegram DM → group → cron trigger)
• Cross-topic calls (Telegram forum topic A → topic B)
• Subagent-spawn calls (direct user → spawned as subagent)
• OpenClaw version upgrades (unless the tool contract meaningfully changes)

Actual behavior

Fingerprint mismatch → sidecar cleared → thread/start issued → new Codex thread with zero history. No user-visible signal except the "forgotten context" symptom. Only log trace is debug-level:

codex app-server dynamic tool catalog changed; starting a new thread

Relevant code

• extensions/codex/src/app-server/thread-lifecycle.ts — startOrResumeThread, fingerprint check gate
• extensions/codex/src/app-server/thread-lifecycle.ts (same file) — fingerprintDynamicTools:

  function fingerprintDynamicTools(dynamicTools) {
    return JSON.stringify(dynamicTools.map(stabilizeJsonValue));
  }

• openclaw-tools.ts — buildMessageToolDescription: injects Current channel (${currentChannel}) into tool description
• openclaw-tools.ts — buildMessageToolSchema: rebuilds inputSchema per-channel actions set
• pi-tools.ts — createOpenClawCodingTools: filters orchestration tools when isSubagentSessionKey(sessionKey) is true

Proposed fix

Option A (minimal, probably sufficient): stabilize fingerprint

Fingerprint only what Codex app-server actually validates across thread/start → thread/resume: tool names + structural shape (property names + required set). Exclude description text and runtime-populated enum values.

function fingerprintDynamicTools(dynamicTools) {
  return JSON.stringify(
    dynamicTools
      .map(t => ({
        name: t.name,
        shape: extractSchemaShape(t.inputSchema),
      }))
      .toSorted((a, b) => a.name.localeCompare(b.name))
  );
}

function extractSchemaShape(schema) {
  // Recursively strip description, default, enum; keep type, required, properties keys, items shape
  ...
}

Semantics: the sidecar should only be invalidated when Codex app-server would genuinely reject thread/resume — i.e., when a previously-known tool is gone, or a tool's argument structure changed in a way the app-server cares about.

Option B (stronger but heavier): compact-on-invalidation

When fingerprint mismatch is detected, call thread/compact/start on the existing thread and seed the resulting summary as developerInstructions on the new thread. This preserves conversational memory across invalidations.

Option C (complementary): scrub runtime context from tool descriptions

buildMessageToolDescription should not bake Current channel (...) into the description. Move runtime context into the system prompt's Environment block (where it already belongs), keeping tool descriptions stable across channels/triggers.

Impact assessment

• Any production agent serving multiple surfaces (DM + group + cron + webhook) cannot reliably use Codex harness today — thread churn is deterministic and frequent
• Teams currently dismissing Codex harness despite GPT-5.x benchmark advantages because continuity is unworkable
• Low-risk fix: Option A alone unblocks most workloads; A+C is ideal

Additional context

• Happy to provide the full 3-sidecar corpus as JSON attachments
• Happy to test a patch on the reporter's setup before merge