hive-keep-codex-fast

Hive Keep Codex Fast

Operator-owned fork of keep-codex-fast.

Default stance: inspect first, preserve continuity, never delete permanently.

Start Here

Resolve this skill directory, then run the script from it.

Read-only report:

python3 scripts/keep_codex_fast.py

Report with raw diagnostics only when the operator asks:

python3 scripts/keep_codex_fast.py --details

Fixture/sandbox path:

python3 scripts/keep_codex_fast.py --codex-home <fixture-or-explicit-codex-home>

Modes

• Runbook: prints a safe off-Codex command sequence, no writes.
• Report: default, privacy-safe, no writes.
• Backup-only: creates backups, does not archive/move/repair.
• Apply: backup-first, archive-only, blocked while Codex holds the state database unless explicitly waiting.
• Metadata repair only: the legacy combined repair path. Backup-first, no archive/prune/move/rotate. Do not reuse it for title-only repair; that follows the stricter contract in references/thread-title-repair-spec.md.
• Title repair dry-run: --repair-thread-titles-dry-run classifies thread titles only, writes no files, and is separate from legacy metadata repair.
• Title repair apply: after dry-run review, --repair-thread-titles-apply --confirm-thread-title-repair APPLY_THREAD_TITLE_REPAIR repairs only eligible threads.title values from safe existing session_index.jsonl.thread_name values. It does not rewrite preview, first_user_message, transcripts, or session_index.jsonl.
• Legacy combined repair: --apply --repair-thread-metadata-bloat also runs normal apply and may compact first_user_message for compatibility.

Before any non-report mode, read references/operation-policy.md.

Report Should Summarize

• active and archived session size
• largest active sessions
• thread title/preview metadata bloat
• stale worktree candidates
• local log size
• bad path/config drift candidates
• top Node/dev/Codex-related processes

Default output must not reveal raw thread IDs, chat titles, private paths, or process args. Use --details only for deliberate diagnosis.

Hard Rules

• Do not run --apply until the operator explicitly approves that exact run.
• Do not archive active repo chats until continuity is preserved or waived.
• Do not treat metadata repair as normal maintenance.
• Do not use legacy metadata repair for title-only repair. Follow references/thread-title-repair-spec.md; legacy first-message compaction requires an explicit operator request and stays on the compatibility path.
• For title-only repair, run --repair-thread-titles-dry-run first and inspect thread_title_dry_run_needs_human before any apply.
• Do not create recurring mutating automation.
• Do not modify memories, skills, plugins, automations, credentials, or API keys.

Operational Guardrails

After report, summarize findings and explicitly say no mutations were made.

Before apply, read references/operation-policy.md, ask the operator to close Codex or approve bounded waiting, and confirm handoff status for important active repo chats.

If the operator is currently using Codex, generate a runbook instead of trying to run mutating maintenance inside Codex:

python3 scripts/keep_codex_fast.py --runbook --codex-home <codex-home>

After apply, run report again and explain backup/archive/restore locations without exposing private backup contents.

References

• references/operation-policy.md: modes, safety gates, apply/backup rules.
• references/thread-metadata-bloat.md: metadata bloat diagnosis and legacy repair context.
• references/thread-title-repair-spec.md: accepted dry-run/apply contract for preserving manual thread titles.
• references/thread-title-repair-prd.md: PRD for the title repair feature before implementation/issues.
• references/handoff-template.md: handoff content before archiving important chats.
• references/evaluation-scenarios.md: pressure tests before install/deploy.
• references/upstream-license.txt: copied upstream license text.