feat: triple-layer post-compaction context enforcement by irchelper · Pull Request #18049 · openclaw/openclaw

irchelper · 2026-02-16T12:41:05Z

After session compaction, agents often skip their configured startup sequence because the compaction summary creates a false 'I already know everything' illusion.

This PR implements triple-layer enforcement to ensure post-compaction recovery:

Layer 1: Summary Append

Append critical workspace rules (AGENTS.md Session Startup + Red Lines) into the compaction summary via <workspace-critical-rules> tags. 100% reliable — rules are physically in the prompt.

Layer 2: System Event

Inject workspace context as a system event for the next agent turn after compaction. Explicit reminder to execute startup sequence.

Layer 3: Post-Turn Audit

After the agent's first post-compaction turn, audit whether required files were actually read. If not, inject a one-time warning listing missing files. Max 1 retry to prevent loops.

Changes

src/agents/pi-extensions/compaction-safeguard.ts — Layer 1: append workspace context to summary
src/auto-reply/reply/agent-runner.ts — Layer 2 + 3: system event injection + audit integration
src/auto-reply/reply/post-compaction-context.ts — reads and formats AGENTS.md sections
src/auto-reply/reply/post-compaction-audit.ts — audits post-compaction file reads
src/auto-reply/reply/post-compaction-context.test.ts — tests for context extraction
src/auto-reply/reply/post-compaction-audit.test.ts — tests for read audit

Design decisions

Reuses existing enqueueSystemEvent() infrastructure
Only activates if workspace has AGENTS.md
extractSections handles H2/H3, case-insensitive, code-block aware
Truncation limits prevent token explosion / Yo-Yo re-compaction
Silent failure on errors (best-effort)
Audit fires at most once per compaction (no infinite loops)
Cross-platform path normalization (forward-slash unification) for Windows compatibility in audit layer

Fixes #18023

…law#18023)

…e blocks)

- Add readWorkspaceContextForSummary() to extract Session Startup + Red Lines from AGENTS.md - Inject workspace context into compaction summary (limited to 2000 chars) - Export extractSections() from post-compaction-context.ts for reuse - Ensures compaction summary includes core rules needed for recovery Part 1 of post-compaction context injection feature.

Windows path.relative() produces backslashes (e.g., memory\2026-02-16.md) which fail to match RegExp patterns using forward slashes. Normalize relative paths to forward slashes before RegExp matching using rel.split(path.sep).join('/'). Fixes 4 test failures on Windows CI.

steipete · 2026-02-17T02:18:29Z

acciedentially merged this but it's problematic:

High – wrong workspace root used for post-compaction context/audit
- src/auto-reply/reply/agent-runner.ts:564 reads workspaceDir =
  process.cwd() before enqueuing context.
- src/auto-reply/reply/agent-runner.ts:599 reuses process.cwd()
  for layer-3 audit.
- followupRun.run.workspaceDir is already passed through (src/
  auto-reply/reply/get-reply-run.ts:413), so this should be
  authoritative.
- Why this is a problem: process.cwd() is process-global and does
  not follow per-session workspace context. In multi-workspace or
  non-default-cwd scenarios, it can point at the wrong repo. That
  means:
  - layer-1/2 context injection can miss the correct AGENTS.md,
    and
  - layer-3 audit can validate reads against the wrong
    workspace.
- Impact: enforcement can silently not run for the intended
  workspace.
Medium – post-compaction audit can pass/fail on stale transcript
data
- src/auto-reply/reply/post-compaction-audit.ts:50-62 reads the
  last 100 lines of the entire session transcript and scans all
  assistant read tool calls.
- src/agents/pi-extensions/compaction-safeguard.ts compaction
  occurs by summarization, not by truncating transcript history.
- Why this is a problem: the “first turn after compaction” check
  is meant to validate startup reads in the turn immediately after
  reset, but current logic can include pre-compaction reads still
  present in transcript history (false pass), and can also miss
  late/older reads if windowing shifts (false fail/noise).
- Impact: weak enforcement and inconsistent warnings in real
  sessions.

will revert

irchelper · 2026-02-17T05:30:13Z

@steipete Hey, thanks for the detailed review on the process.cwd() and stale transcript issues — you were absolutely right about both problems.

I actually spotted these bugs shortly after submitting the PR, but I was hit by AI model rate limits at the time and couldn't get my agent system to work on fixes (I work on a strict "AI-only, zero manual coding" principle 😅).

When the rate limits finally cleared, I did a full audit:

Bug 2 (stale transcript): Already fixed upstream — getPostCompactionContext() now anchors from compaction summary instead of the fixed slice(-100) window. ✅
Bug 1 (process.cwd()): Still present in agent-runner.ts lines 564 and 599. The upstream compaction safeguard (compact.ts) uses the correct effectiveWorkspace path, so the main compaction flow is unaffected — but the post-compaction context injection in agent-runner.ts still relies on process.cwd(), which is wrong for multi-workspace / non-default-cwd scenarios.

I'm working on a fix for Bug 1 now and will submit a new PR shortly.

On a personal note — this was my very first time using GitHub to submit issues and PRs, so I'm still learning. It's also Chinese New Year right now, which added to the delay.

Since then I've set up a full HA multi-model setup with four providers (Claude, Gemini, GPT-5.x, MiniMax) with task-type routing, so I should be able to keep contributing going forward. Thank you for open-sourcing OpenClaw — it got me back into programming after 20+ years away from code. Turns out directing AI workers is a pretty fun way to build things!

irchelper · 2026-02-17T05:43:59Z

Quick follow-up: I've submitted #18914 to fix Bug 1 (process.cwd() → followupRun.run.workspaceDir).

I'm aware that upstream's compaction safeguard in compact.ts already uses the correct effectiveWorkspace path via resolveBootstrapContextForRun(), so the main compaction flow is unaffected by this bug. But the post-compaction context injection code in agent-runner.ts still had the incorrect process.cwd() calls, and I wanted to close the loop properly rather than leave known bugs in the codebase.

Add notifyOnStart and notifyOnStartText fields to AgentCompactionConfig so users receive a message the moment auto-compaction begins, resolving the UX problem where the interface appears frozen or shows red errors with no context. Changes: - src/config/types.agent-defaults.ts: add notifyOnStart (boolean) and notifyOnStartText (string) to AgentCompactionConfig - src/config/zod-schema.agent-defaults.ts: add corresponding Zod fields inside the .strict() compaction object - src/auto-reply/reply/agent-runner-execution.ts: add onCompactionStart optional callback to runAgentTurnWithFallback params; call it when the Pi runtime fires compaction:phase=start - src/auto-reply/reply/agent-runner.ts: build and pass onCompactionStart callback that reads cfg.agents.defaults.compaction.notifyOnStart and delivers via opts.onBlockReply when enabled - src/config/config.compaction-settings.test.ts: 5 new tests covering true/false/undefined and schema rejection of invalid types - src/auto-reply/reply/agent-runner.compaction-notify.test.ts: 5 new agent-runner integration tests (default text, custom text, false, undefined, end-only event) - docs/concepts/compaction.md: document the new config fields - docs/reference/session-management-compaction.md: add implementation reference section Extends openclaw#18049 (triple-layer post-compaction context enforcement). Default: notifyOnStart is false (opt-in, no behaviour change for existing users).

feat: inject post-compaction workspace context as system event (openc…

8ce59cc

…law#18023)

openclaw-barnacle bot added the size: M label Feb 16, 2026

康熙 added 3 commits February 16, 2026 20:44

fix: format post-compaction-context test file

b4b7316

fix: improve section extraction robustness (case-insensitive, H3, cod…

6ba1d9a

…e blocks)

openclaw-barnacle bot added the agents Agent runtime and tooling label Feb 16, 2026

irchelper changed the title ~~feat: inject post-compaction workspace context as system event~~ feat: dual-layer post-compaction context enforcement (summary append + system event) Feb 16, 2026

fix: format compaction-safeguard.ts with oxfmt

3917f3c

irchelper changed the title ~~feat: dual-layer post-compaction context enforcement (summary append + system event)~~ feat: triple-layer post-compaction context enforcement Feb 16, 2026

feat: add post-compaction read audit (Layer 3)

fc10867

openclaw-barnacle bot added size: L and removed size: M labels Feb 16, 2026

steipete merged commit 65a1787 into openclaw:main Feb 16, 2026
23 checks passed

irchelper mentioned this pull request Feb 16, 2026

Post-compaction session loses injected project context (AGENTS.md, SOUL.md, etc.) #17727

Closed

irchelper deleted the feat/post-compaction-context-inject branch February 17, 2026 00:43

irchelper mentioned this pull request Feb 17, 2026

fix: use per-session workspaceDir instead of process.cwd() for post-compaction context #18914

Open

This was referenced Feb 19, 2026

feat(compaction): notify user on compaction start irchelper/openclaw#1

Closed

feat(compaction): notify on start + idle-triggered proactive compaction (extends #18049) #21117

Open

kojira mentioned this pull request Feb 26, 2026

[Security] post-compaction-audit.ts hardcodes WORKFLOW_AUTO.md — viable attack vector for persistent prompt injection #27697

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: triple-layer post-compaction context enforcement#18049

feat: triple-layer post-compaction context enforcement#18049
steipete merged 7 commits intoopenclaw:mainfrom
irchelper:feat/post-compaction-context-inject

irchelper commented Feb 16, 2026 •

edited

Loading

Uh oh!

Uh oh!

steipete commented Feb 17, 2026

Uh oh!

irchelper commented Feb 17, 2026 •

edited

Loading

Uh oh!

irchelper commented Feb 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

irchelper commented Feb 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Layer 1: Summary Append

Layer 2: System Event

Layer 3: Post-Turn Audit

Changes

Design decisions

Uh oh!

Uh oh!

steipete commented Feb 17, 2026

Uh oh!

irchelper commented Feb 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

irchelper commented Feb 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

irchelper commented Feb 16, 2026 •

edited

Loading

irchelper commented Feb 17, 2026 •

edited

Loading