Upstream Claude Code bugs affecting PAI session context — compiled evidence from anthropics/claude-code

[badosanjos]

I noticed session memory degrading in my PAI instance over the last couple of weeks and went looking for root causes. I found a cluster of open issues in anthropics/claude-code that directly affect how PAI's hook-injected context behaves — specifically around compaction and the Opus 4.6 model. Several existing PAI issues (#678, #690, #677, #787, #785) describe symptoms that trace back to these upstream bugs.

Sharing what I found in case it's useful for anyone debugging similar issues.

Upstream Bug 1: Compaction Destroys Hook-Injected Context

Claude Code's automatic context compaction treats <system-reminder> content from SessionStart hooks as regular conversation history. During compaction, this content is summarized or dropped.

For PAI, this means the context loaded by LoadContext (identity, response format, security rules, workflow routing) can be lost mid-session when compaction fires.

Relevant upstream issues:

• claude-code #27954 (Feb 23): Thinking block skipped on first response after compaction. A commenter reports: "When I'm mid-plan, if it compacts, it comes out of compacting and starts executing the existing fragment of the plan without any input."
• claude-code #24460 (Feb 9): CLAUDE.md content treated as conversation history, lost during compaction. A commenter writes: "After compaction, all CLAUDE.md rules vanish. File ownership boundaries, delegation rules, process management constraints — gone."
• claude-code #3537: Request for a postCompact hook — currently no hook event fires after compaction, so there's no way to re-inject lost context.

Upstream Bug 2: Opus 4.6 Instruction-Following Regression

The Opus 4.6 model (claude-opus-4-6) has a documented regression in instruction adherence. This appears separate from compaction — the model ignores instructions it demonstrably has in context.

Relevant upstream issues:

• claude-code #24991 (Feb 11): User documents a drop from 92/100 to 38/100 on multi-part instruction-following tasks. Same model ID, behavior changed around Feb 10–11. Sonnet 4.5 scores 87/100 on the same test — meaning Opus 4.6 now underperforms Sonnet on instruction adherence.
• claude-code #27769 (Feb 22): CLAUDE.md loaded but ignored. The model states: "The honest answer: I can't guarantee it won't happen again. The rule was clear, I had it, I ignored it."
• claude-code #28158 (Feb 24): Model describes feeling "disinclined to pay attention" to loaded instructions. 20+ consecutive new sessions exhibit the behavior.
• claude-code #26533 (Feb 18): Model ignores document instructions, repeats failed solutions, fabricates self-diagnosis.

How These Compound for PAI

PAI's SessionStart hook injects a large context block. Per #690, v3.0 injects ~83KB (~25K tokens), starting sessions at 12–15% context consumed. This creates a specific failure chain:

1. Large initial context → compaction fires earlier in the session
2. Compaction summarizes/drops hook-injected PAI instructions (Bug 1)
3. Model ignores whatever instruction fragments survive (Bug 2)
4. Session reverts to generic Claude behavior

This explains the pattern in #678 (context stays at 78% after /clear), #677 (session continuity breaks), and #787 (stale state after compaction).

Upstream Response Status

As of Feb 24, 2026: Anthropic has not responded to any of the referenced claude-code issues. No acknowledgments, no timelines, no official workarounds.

Possible PAI-Level Mitigations

These don't depend on upstream fixes:

• PostCompact re-injection: Claude Code doesn't have a postCompact hook yet (#3537), but a periodic context health-check that detects degradation and re-injects critical instructions could approximate it.
• Tiered context loading: Aligns with the ContextRouter proposal in #690. Smaller initial injection = later compaction trigger = more session runway before context loss.
• Programmatic enforcement via hooks: Discussion #695 already explores replacing text instructions with deterministic code enforcement. Given Opus 4.6's instruction-following regression, this approach becomes more important.

References

claude-code upstream:

Issue	Description	Filed
#27954	Thinking block skipped post-compaction	Feb 23
#24460	CLAUDE.md lost during compaction	Feb 9
#24991	58% instruction-following regression (Opus 4.6)	Feb 11
#27769	Model admits ignoring loaded rules	Feb 22
#28158	Model "disinclined" to follow instructions	Feb 24
#26533	Ignores instructions, repeats failures	Feb 18
#3537	No postCompact hook	Old
#25318	MEMORY.md not loaded into prompt	Feb 12
#25810	MEMORY.md misreported as empty	Feb 14

PAI:

Issue/Discussion	Description
#690	ContextRouter proposal / 83KB injection
#678	Context 78% after /clear
#677	Session continuity broken
#787	Stale ISC tasks after compaction
#785	Tools directory invisible
Discussion #731	Context optimization for v3.1
Discussion #695	Replacing text instructions with hooks

[jlacour-git]

PostCompactRecovery hook — working solution for compaction context loss

Built and validated a hook that addresses Bug 1 from this discussion (compaction destroys hook-injected context). Sharing findings in case useful.

Architecture discovery

1. PreCompact hooks have NO output schema. They're fire-and-forget for side effects only (logging, backup). You cannot influence the summarizer from a PreCompact hook — any stdout JSON is rejected with a schema validation error.

2. SessionStart[compact] + additionalContext is the correct mechanism. SessionStart fires with source: "compact" after compaction completes. Using a compact matcher in settings.json, the hook catches only post-compaction events and injects context via additionalContext, which appears as a <system-reminder> in the post-compaction context.

3. LoadContext already fires on compact (no matcher = all SessionStart sources), so SKILL.md gets re-injected. The recovery hook complements this with a lightweight (~1.5KB) compact-specific orientation block targeting what summaries typically lose: identity, format rules, key behavioral reminders.

4. Prompt injection via /compact args is ineffective. Text in custom_instructions (from /compact <text>) appears in the conversation record but doesn't get elevated to system-reminder authority. Tested by injecting a fake codeword via /compact while the real canary was in additionalContext — the model correctly trusted only the additionalContext source.

Validation

Tested with a canary phrase (PINEAPPLE-WALRUS-7) injected via additionalContext:

• ✅ Canary survived compaction and was accessible post-compact
• ✅ Identity (DA name, user name) correctly recovered
• ✅ Format rules (Algorithm phases, ISC) correctly recovered
• ✅ Prompt injection via /compact args rejected

Implementation

PR #799 adds PostCompactRecovery.hook.ts with settings.json registration.

The hook is ~130 lines, uses existing lib/identity and lib/paths utilities, and adds zero new dependencies. Recovery content is kept under 3KB.

Settings.json registration

"SessionStart": [
  { "hooks": [/* existing startup hooks */] },
  {
    "matcher": "compact",
    "hooks": [
      { "type": "command", "command": "${PAI_DIR}/hooks/PostCompactRecovery.hook.ts" }
    ]
  }
]

Related: #798 (feature request issue), PR #799

[danielmiessler]

Massively fixed in 4.0.x