Ca Prune Skill for Claude

/ca-prune — session transcript pruner

Long sessions die when their semantic history fills the context window with bulk: oversized tool

outputs, thinking blocks, MCP/shell noise, stale file reads, and host-specific sidecars. The shared

policy trims that bulk at safe quiescence boundaries while each host codec preserves its native

integrity model and the K most recent tool-bearing turns verbatim. Claude keeps byte-safe JSONL

serialization; Pi uses semantic entries and returns a native custom-compaction result.

Gains land at the host's next native resume/restart or compaction boundary, not by mutating the

running session's active file. Pi-native compaction is event-driven and never rewrites an active

Pi session file.

Argument

$ARGUMENTS is one of:

• status (default) — report cumulative reduction and service state for this session from

~/.codearbiter/prune-state.json, and whether CODEARBITER_PRUNE is on/dry/off. When

the service has run in dry mode, every would-be prune is also recorded — one JSONL row per

decision, across all sessions — to the shared data-collection log

~/.codearbiter/metrics/prune-dry.jsonl (override with CODEARBITER_PRUNE_METRICS). That log is

the evidence base for the dry→on decision: a clean record (every row verdict: dry-run,

validation_errors: 0) over a representative set of sessions is the signal that enabling is safe.

Read context_bytes_freed / context_est_tokens_freed for model-context benefit and

file_bytes_freed / file_pct for disk and resume-parse benefit. sidecar-collapse is explicitly

file-only; it must not count toward the context-benefit or cold-cache decision.

The legacy freed_bytes, pct, and est_tokens_before / est_tokens_after fields remain

whole-file compatibility aliases; do not use them as model-context evidence.

• dry — create a read-only semantic plan (or analyze a scratch copy where the host exposes a

serialized transcript); present the per-strategy reduction table with every strategy labeled

context or file-only. Never writes the active session.

• run <path> — prune the target with --execute. Targets a **copy or an old/inactive

transcript only** — the tool refuses an active or recently-modified target by construction.

• audit <path> — read-only integrity report: line-parse, uuid chain, tool-pair coverage,

condensation markers.

• on / off — guidance on enabling or disabling the after-each-turn service.

Flow

• status / dry / audit — run the backing tool and present its output verbatim:

   python3 "<plugin-root>/hooks/prune-transcript.py" <subcommand> [<path>] || python "<plugin-root>/hooks/prune-transcript.py" <subcommand> [<path>]

For serialized hosts, dry analyzes <path>.copy.jsonl; Pi active sessions use the native

semantic planner and return a custom compaction result without session-file writes.

• run — confirm the path is a copy or an inactive session, then:

   python3 "<plugin-root>/hooks/prune-transcript.py" <path> --execute [--tier T] || python "<plugin-root>/hooks/prune-transcript.py" <path> --execute [--tier T]

Present the per-strategy reduction report; follow with audit on the result.

• on/off — explain the after-each-turn service: UserPromptSubmit and PreCompact hooks

prune at safe quiescence points, always exit 0, and never block the prompt. Tiers: gentle

(sidecar + oversize clamp), standard (+ reasoning fold, aged/MCP/shell), aggressive

(+ stale-read, reminder dedup, image evict). Config via CODEARBITER_PRUNE (off|dry|on,

ships off), CODEARBITER_PRUNE_TIER, CODEARBITER_PRUNE_KEEP_RECENT (the K most recent

tool turns kept verbatim — each turn is an assistant tool_use plus its results),

CODEARBITER_PRUNE_MAXBYTES. Enabling is the user's explicit choice — never set it unbidden.

In dry mode the service writes no transcript but appends each would-be prune to

~/.codearbiter/metrics/prune-dry.jsonl (path override: CODEARBITER_PRUNE_METRICS) for

data collection; in on mode the executed prunes are recorded in ~/.codearbiter/prune.log.

If a prior service-mode prune was killed mid-write, the next run self-heals the transcript

from the newest backup in ~/.codearbiter/prune-backups/ before doing anything else.

Cold-miss nudge [Feature Forge — preview]: when CODEARBITER_PRUNE is on, an

optional submit-time speed bump warns once before a cold cache re-cache lands on bloated

context. Enable with CODEARBITER_PRUNE_NUDGE=on (default off). When all arming conditions

hold (idle ≥ CODEARBITER_PRUNE_NUDGE_IDLE_SECS, default 240 s; estimated model-context

tokens freed ≥

CODEARBITER_PRUNE_NUDGE_MIN_TOKENS, default 80 000), the hook blocks the submit once with

an advisory on stderr and returns exit code 2. File-only sidecar reduction never arms it. The

advisory names the approximate avoidable context-token count and the host-native actions that

move the re-cache to pruned context:

native compaction or a normal exit + resume/restart. Resubmitting immediately proceeds. The block fires at most

once per cold window; a subsequent warm submit (idle < floor) resets the window so the next

genuine cold stretch re-arms. The gate is strictly opt-in, never fires in dry/off mode,

and fails open on any error — a pruner fault will never block the session.

When NOT to use

• Context bar nowhere near compaction — the most recent turns are protected anyway.
• Install health → /ca-doctor.
• Project progress → /ca-status.

Hard gate

• MUST NOT run --execute against the live session's transcript — the tool refuses a

recently-modified file by construction; manual run targets copies or old sessions only.

• MUST surface the native-boundary-only gains limitation whenever a user expects an immediate

active-file context drop.

• MUST NOT enable the service (CODEARBITER_PRUNE=on) on behalf of the user — explain and let

them decide.