Run Space Reclamation

The current command surface still lists `prune` and `purge`, and dispatch routes them through the same bare first-argument rule as other run-management verbs. Their request parser resolves a single run name or an explicit all-runs scope, accepts dry-run and confirmation bypass flags, and rejects missing or conflicting scopes before planning.

The candidate planner still separates requested runs, eligible runs, and skipped runs. It protects active and live-owned runs, applies age filters conservatively, inventories transcripts, stage logs, and usage sidecars, and skips tracked artifacts with local modifications. Those decisions feed both preview and apply paths.

The reporting code now renders an artifact action plan before any prompt or write. It shows scope, action meaning, skipped runs, per-artifact actions or skip reasons, and totals from the same plan the applier receives. Later transcript work replaced the early placeholders with transform-backed accounting rather than changing the report surface.

The operation path still treats dry runs, empty selections, declined prompts, confirmed runs, sidecar creation, preservation failures, and action failures as explicit outcomes. Missing usage sidecars are derived before prune or purge proceeds, and a preservation failure blocks that run from deletion or rewrite. The outcome renderer reports those cases separately.

The CLI reference still documents the reclamation family beside other run-management verbs. It distinguishes reclamation from teardown, explains explicit scope and safety behavior, and keeps compatibility expectations around destructive commands in one place. Some original integration fixture files have since been reorganized, but the command documentation remains current.

The current artifact action plan still enumerates transcripts, stage logs, and usage sidecars with an action or skip reason per row. Report rendering uses that plan for dry runs and confirmation previews, and purge and prune both consume the same artifact rows. The earlier aggregate-only shape has been replaced by concrete artifact reporting.

`prune_transcript_bytes` remains the core transcript transform. It replaces repository-sized file payloads with markers, truncates long command output while keeping readable head and tail text, preserves ordinary JSON lines, and recognizes already-pruned transcripts as no-work. The transform supplies the byte accounting used by prune previews.

Confirmed prune now rewrites only the transcript rows named by the plan. Before replacing a file it re-runs the transform and refuses the rewrite if the current bytes no longer match the preview, then writes atomically. Tests still cover idempotent pruning and retained command identity, status, exit information, usage data, and final result classification.

Confirmed purge now deletes only planned detailed transcript and stage-log artifacts after preservation has run. The implementation keeps usage sidecars and durable run materials, and it checks the planned byte state again before deletion. The CLI docs still state that purge is not run removal and that the run remains listable at summary level.

The CLI reference, versioning guide, replay corpus runbook, and screencast runbook still describe pruned and purged runs. They explain what transcript-dependent workflows lose, what replay can still use, and why retained markers and missing detailed logs are compatibility-relevant states. That documentation remains the operator-facing bridge between cleanup and downstream tooling.

Purge planning now includes detailed stage logs as reclaimable artifacts and preserves usage sidecars as durable facts. Missing sidecars are accounted for even when the apply step must create them before deletion. The report and outcome paths still show these preservation and deletion decisions explicitly.

The tests now exercise real Claude and Codex transcript semantics after pruning, including usage extraction and final-result parsing. Small fixture files for opencode and Gemini remain in the tree to capture the bounded-payload finding. This means prune behavior is still checked against the harness differences the plan called out.

`WorktreeReclamationCandidate` still carries repository identity, run identity, checkout path, branch, age eligibility, live owner, Git-known state, merge safety, local risk, byte measurements, selected action, and skip reasons. Candidate planning also includes orphaned directories under the repository worktree prefix, so cleanup can see historical leftovers outside the run registry.

Confirmed worktree reclamation now removes eligible orphaned or fully merged disposable checkouts and removes known merged run branches through Git. It rechecks risk before deletion, refuses widened bare invocations, supports age filters, and reports byte mismatches or changed safety state as failures to rerun from a fresh plan. Dirty, unmerged, live, missing-branch, and unknown-safety checkouts stay protected by default.

`build-directories` now removes rebuildable output such as Rust target directories while leaving the checkout itself in place. It uses the same report, dry-run, confirmation, all-runs scope, and age-filter machinery as worktree removal. Candidates with no build output are reported separately, and live worktrees remain protected.

The lifecycle tests now cover normal terminal cleanup removing the disposable checkout and branch, and failed or handoff paths retaining state for resume or by-hand repair. `src/engine/cleanup.rs` still documents teardown as botched-run removal, distinct from historical reclamation. The CLI docs direct historical leftover checkouts to the worktree reclamation report.

The command reference now describes every reclamation choice in one place: doing nothing, pruning transcripts, purging detailed logs, removing stale worktrees, and cleaning rebuildable output. The versioning guide records the command-surface and artifact-format implications, while corpus and screencast docs state what remains possible after logs are pruned or purged.