agentsclimarketplace

Persist

Skill broomva/skills/skills/orchestration/persist

bstack P12 — Persistent Loop Discipline. Cross-context restart loop where state lives in the filesystem (PROMPT.md + git tree + state.jsonl), not in the conversation. Each iteration spawns a fresh agent context. Solves the 'context rot' failure mode where long-horizon agentic work (>1h, METR's 80%-reliability ceiling) degrades silently past ~100K tokens. Validation backpressure comes from compilers/tests/linters, not model self-grading. Use persist when: (1) starting work that may span hours and exceed the model's reliability horizon, (2) detecting context drift mid-session (token usage past 100K, repeated failed iterations on the same fix), (3) coordinating long-horizon work that needs to survive crashes / context exhaustion, (4) running parallel work streams (one persist loop per worktree, composed with bstack P5). Triggers on 'persist', 'long-horizon loop', 'context restart', 'fresh-context iteration', 'P12', 'Ralph loop', 'filesystem-state loop'.From its SKILL.md

Install
npx -y skills add broomva/skills --skill persist

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

6.1 KB, ~1.3k tokens by cl100k_base, as published. Nobody here has run it

persist — bstack P12 Persistent Loop Discipline

Cross-context restart loop. State in filesystem, not conversation.

The defining moves:

  1. The agent writes a goal + state snapshot to PROMPT.md
  2. persist iterate PROMPT.md spawns a fresh agent context per iteration
  3. State persists in the filesystem (PROMPT.md + git tree + state.jsonl)
  4. Validation backpressure: compilers/tests/linters, not model self-grading
  5. Loop exits when success_condition fires OR budget exhausted OR user interrupts

Why this exists

METR's Time Horizon 1.1 puts the 80%-reliability deployable horizon at ~1 hour on Opus 4.6. Above that, model coherence degrades silently — context rot past ~100K tokens (the Dumb Zone). In-context loops (ReAct/TAO) fail because they share the rotting context window. Persist solves this by restarting the context every iteration while keeping state in the filesystem.

When to invoke

The reflexive trigger rule (full text in workspace AGENTS.md §P12):

  1. Before starting any work that may exceed ~1h of unsupervised agent time — write PROMPT.md, decide budget, pick success condition, call persist iterate.
  2. When token usage in the current session crosses ~100K — restart instead of continuing in the rotted context.
  3. When the same fix has been attempted ≥3 times without convergence — stop the in-context loop; write the diff history to PROMPT.md and start fresh.
  4. When orchestrating long-horizon work — default to persist with periodic checkpoints; compose with P5 worktrees for parallel persist loops.

CLI

persist iterate PROMPT.md \
  --max-iterations 50 \
  --max-wall-clock 14400 \
  --success-condition "grep:DONE:STATUS" \
  --agent-cmd "claude -p '{}'"

persist status                  # show open loops
persist status --json           # machine-readable
persist abandon <loop-id>       # terminal: mark ABANDONED, free slot
persist doctor                  # health-check (state dir, git available)
persist conformance             # run test battery

The {} token in --agent-cmd is replaced with the prompt file's contents. Default agent is claude -p '{}'. Codex: --agent-cmd "codex {}". Gemini CLI: --agent-cmd "gemini -p '{}'".

Success conditions

Three forms:

  • exit-code-0 — last agent invocation returned 0
  • file-exists:PATH — agent writes a sentinel file when done
  • grep:PATTERN:FILE — agent writes a status line that matches PATTERN

The agent is responsible for updating PROMPT.md or writing the sentinel file at the end of each iteration. The script doesn't try to interpret agent output — that's the backpressure must come from external signals invariant.

State machine

SPAWNED ──→ ITERATING ──→ ITERATING ──→ ... ──→ SUCCESS (terminal)
                  │                          ╲
                  ↓                           ↘ BUDGET_EXHAUSTED (terminal)
              PAUSED ──→ ITERATING            
                  │
                  ↓
              ABANDONED (terminal)

State events append to ~/.config/broomva/persist/state.jsonl (JSONL append-only with flock).

Composition with bstack

primitivecomposes via
P5 Parallel Agentsrun N persist loops, one per git worktree
P7 CI Watchereach iteration's pushed PR uses p9 watch for productive-wait
P10 Worktree Hygieneclean tree before iteration; janitor after each merge
P11 Empirical Feedbackper-iteration validation; persist's success_condition is multi-modal evidence
P6 Bookkeepingpersist loops produce graph-relevant material → bookkeeping replay between loops

Invariants

  • State lives in the filesystem. Each iteration starts from PROMPT.md content, not conversation history.
  • Validation backpressure is external. Don't ask the agent "are you done?" — check exit codes, file presence, or status pattern.
  • Budget bounds must be honored. Default 50 iterations / 4h wall-clock. The 4h default matches METR's 80%-horizon ceiling.
  • State.jsonl is append-only. Loop terminations are terminal — no resurrection. To restart, spawn a new loop with a new ID.
  • Each iteration is a fresh process. persist calls the agent CLI in a subprocess; agent context never persists between iterations except via filesystem state.

See also

Background

Pattern popularized by Geoffrey Huntley as "everything is a ralph loop" (Jan 2026). Anthropic shipped a ralph-wiggum plugin; OpenAI shipped /goal in Codex CLI 0.128.0. bstack's P12 is the same mechanism with non-anthropomorphized naming and explicit composition with the rest of the bstack contract.

What ships with it: 4 files

25.3 KB alongside SKILL.md, 3 of them executable

scripts/

Keep looking

Skills are one crate of 326,144. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.