openclaw - 💡(How to fix) Fix [Bug]: post-compaction context injection uses process.cwd() instead of run.workspaceDir (agent-runner.runtime line 3312) [1 participants]

Summary

readPostCompactionContext is called with process.cwd() instead of the run's configured workspaceDir inside the auto-compaction post-hook for follow-up runs (agent-runner.runtime-CH0aH7T6.js line 3312). For agents that run via a CLI provider subprocess (e.g. openai-codex), process.cwd() inside the subprocess is typically $HOME, so the gateway repeatedly tries to read $HOME/AGENTS.md and logs Warning: Could not read $HOME/AGENTS.md: Error: EISDIR... (or ENOENT depending on what happens to be there).

Location

/opt/homebrew/lib/node_modules/openclaw/dist/agent-runner.runtime-CH0aH7T6.js (v2026.4.14), line 3312:

if (sessionKey) readPostCompactionContext(process.cwd(), cfg).then((contextContent) => {
    if (contextContent) enqueueSystemEvent(contextContent, { sessionKey });
}).catch(() => {});

A few hundred lines earlier, line 1409 in the same file does the correct thing:

const refreshPrompt = await readPostCompactionContext(params.followupRun.run.workspaceDir, params.cfg);

followupRun.run.workspaceDir is in scope at line 3312 too — line 3290 already uses followupRun.run.sessionId. So the two call sites are inconsistent for no clear reason.

Reproduce

1. Configure a default heartbeat in openclaw.json:

   "agents": {
     "defaults": {
       "heartbeat": { "model": "openai-codex/gpt-5.4", "every": "30m" }
     }
   }

2. Let the heartbeat fire a few times. When auto-compaction triggers on the heartbeat follow-up run and the Codex CLI subprocess's cwd is $HOME, the gateway starts logging Warning: Could not read $HOME/AGENTS.md on every occurrence.
3. The problem is especially visible if something happens to exist at $HOME/AGENTS.md (directory or file). If nothing is there, the warning is still emitted (noisy but benign).

Not reproducible with provider runtimes that don't spawn a CLI subprocess with $HOME as cwd (e.g. Anthropic, Gemini direct), because those go through the other code path that uses the configured workspaceDir.

Expected

Post-compaction context injection should load AGENTS.md from the agent's configured workspaceDir — the same file the pre-compaction refresh reads — not from whatever cwd the runner subprocess happens to have.

Suggested fix

One-line change to line 3312:

- if (sessionKey) readPostCompactionContext(process.cwd(), cfg).then((contextContent) => {
+ if (sessionKey) readPostCompactionContext(followupRun.run.workspaceDir ?? process.cwd(), cfg).then((contextContent) => {

The ?? process.cwd() fallback keeps existing behavior as a defensive default if a caller ever omits workspaceDir. Alternatively, match line 1409 exactly and drop the fallback.

Impact

Low severity — the warning is non-fatal because the downstream .catch(() => {}) swallows the error, so no user-facing breakage. But:

• Log noise: ~48 warnings/day per affected agent (every 30 min heartbeat tick)
• Post-compaction context injection does not fire for affected agents when the warning triggers, so the agent loses the intended "re-read Session Startup / Red Lines" nudge after compaction — which is exactly the scenario the feature was designed to cover
• Obscures the useful signal in gateway.err.log when actually investigating a problem

Environment

• OpenClaw 2026.4.14
• macOS 26.3.0 / Apple Silicon
• Heartbeat provider: openai-codex/gpt-5.4
• Agent without explicit workspace override (inherits agents.defaults.workspace)

Related / not-dupes

• #37398 (closed) — "Inject configurable workspace files into post-compaction context" — adjacent feature, not this bug
• #25392 (closed) — "Default AGENTS.md template headings mismatch compaction code" — different bug in the same area
• #25969 (closed) — "support after_compaction internal hooks + followup turn" — the PR that likely introduced this code path

Happy to submit a PR if helpful.