harden-wg-pi-compaction-kick — WorksGood live mirror

Metadata

Status	open
Agent identity	`02e879681e52e0a384106169be043416c4d946e850ab26b2269c57681b52a6e7`
Created	2026-08-05T20:52:42.667747065+00:00
Tags	`pi`, `compaction`, `adversarial`, `validation`

Description

Objective

Independently review, adversarially test, and repair the implemented Pi compaction kick. Treat the implementation task’s tests as claims, not proof. Re-run the original reproducer against the installed binary and inspect the raw Pi NDJSON, durable session journal, watchdog state/journal, lifecycle ledger, wrapper PIDs, and task outcome. Make surgical fixes for any defect found.

Adversarial cases must include duplicate compaction_end and agent_settled, terminal receipt racing permit and delivery, already queued follow-up, two successive threshold compactions, crash after session append but before acknowledgement, child exit during delivery, exact-session file ambiguity, stale process epoch/PID reuse, budget exhaustion, and a normal completed task whose last turn also triggers compaction. Verify that attended Pi/manual /compact behavior is unchanged.

Validation

Original credential-free real-flow reproducer fails on the recorded pre-fix baseline and passes on the installed candidate.
Every adversarial case has a durable exactly-once assertion over action ID, prompt/message, continuation charge, process/session/attempt identity, tool effects, and terminal disposition.
No case can convert silence/compaction/exit into Done or Failed without the existing authoritative receipts/finalization path.
No over-trigger occurs for completed, queued, manual, overflow, failed-compaction, chat, console, or unsafe-effect cases.
Embedded extension bytes are in sync with source and the globally installed binary is rebuilt from the final tree.
cargo fmt --check, cargo clippy, focused/full Rust and plugin tests, all smoke scenarios owned by both implementation and hardening tasks, and cargo install --path . --locked pass; any environmental skip is reported, not counted as pass.

## Objective
Independently review, adversarially test, and repair the implemented Pi compaction kick. Treat the implementation task’s tests as claims, not proof. Re-run the original reproducer against the installed binary and inspect the raw Pi NDJSON, durable session journal, watchdog state/journal, lifecycle ledger, wrapper PIDs, and task outcome. Make surgical fixes for any defect found.

Adversarial cases must include duplicate `compaction_end` and `agent_settled`, terminal receipt racing permit and delivery, already queued follow-up, two successive threshold compactions, crash after session append but before acknowledgement, child exit during delivery, exact-session file ambiguity, stale process epoch/PID reuse, budget exhaustion, and a normal completed task whose last turn also triggers compaction. Verify that attended Pi/manual `/compact` behavior is unchanged.

## Validation
- [ ] Original credential-free real-flow reproducer fails on the recorded pre-fix baseline and passes on the installed candidate.
- [ ] Every adversarial case has a durable exactly-once assertion over action ID, prompt/message, continuation charge, process/session/attempt identity, tool effects, and terminal disposition.
- [ ] No case can convert silence/compaction/exit into Done or Failed without the existing authoritative receipts/finalization path.
- [ ] No over-trigger occurs for completed, queued, manual, overflow, failed-compaction, chat, console, or unsafe-effect cases.
- [ ] Embedded extension bytes are in sync with source and the globally installed binary is rebuilt from the final tree.
- [ ] `cargo fmt --check`, `cargo clippy`, focused/full Rust and plugin tests, all smoke scenarios owned by both implementation and hardening tasks, and `cargo install --path . --locked` pass; any environmental skip is reported, not counted as pass.

Depends on

Required by

open implement-wg-single-attempt-kernel

Log

2026-08-05T20:52:42.618925840+00:00 Task paused
2026-08-05T20:52:42.761701640+00:00 Task published
2026-08-05T20:52:50.219277573+00:00 Lightweight assignment: agent=Careful Programmer (02e87968), exec_mode=full, context_scope=full, reason=The Careful Programmer is the sole and best fit for this correctness-critical WG-wide adversarial validation and surgical repair task, which requires full repository access and modification.