Patterns from PrimeIntellect prime-agent worth adopting in ArcKit

[tractorjuice]

Context

Review of PrimeIntellect-ai/prime-agent — an open-source coding/research agent built around a Recursive Language Model (RLM) runtime (persistent IPython kernel as the only built-in tool) and a Continual Harness (supplemental prompts/memories/skill descriptions refined via small, evidence-backed updates).

Much of it is host-level runtime engineering a Claude Code plugin can't replicate (persistent kernel, programmatic rlm(...) subagent spawning, daemon-backed detachable sessions, agent-to-agent messaging, heartbeats). Claude Code already covers the practical ends of those via background agents, Routines/send_later, and the Monitor tool, which ArcKit's docs already point at. But several patterns transfer directly.

Adoptable patterns, ranked

1. Quality gates with bounded retry for arckit-build (highest value)

Prime-agent's autonomous mode runs configured gate commands before a run may finish; a failed gate feeds its bounded output back to the agent for another attempt, and it skips re-running a gate when the workspace hasn't changed.

The build harness currently halts on failure and asks the user to --resume. Proposal:

• On validation failure, feed the failure output back to the writer subagent for 1–2 bounded retries before halting the wave.
• Skip re-running an unchanged gate — composes naturally with the existing SHA-256 input-hash staleness detection in skills/arckit-build/SKILL.md.
• Adopt their honesty doctrine verbatim in build summaries: "a passed gate checks only what that gate verifies; reaching a limit does not imply task success."

2. Incomplete-build nudge (persistent-goal analog) — cheapest win

Prime-agent's /goal keeps an objective alive across turns until explicitly completed, with token-budget accounting. ArcKit already has the pieces: state.json records build progress and the Stop-hook nudge engine (session-nudge.mjs) already fires traceability suggestions.

Proposal: a new nudge rule — "build for project 001 is incomplete at wave 3 of 6 — suggest --resume" — replicating goal persistence entirely within the existing hook model.

3. Refinement snapshots + rollback for the self-harness loop

ArcKit's autoresearch self-harness (weakness-miner.mjs → harness-proposer.mjs → held-in/held-out harness-validator.mjs) is more rigorous than prime-agent's /refine on acceptance criteria. What prime-agent adds:

• every refinement event records before/after snapshots supporting rollback;
• the base prompt is immutable, refinements are supplemental layered state;
• refinements are session-local by default, promoted to global deliberately.

If the autoresearch flow doesn't already snapshot-and-rollback accepted harness edits, that's the gap — it turns "the optimizer made things worse" from archaeology into a one-command revert.

4. Structured resume-brief format

Prime-agent's compaction summary has a fixed contract: Goal / Constraints & Preferences / Progress (Done · In Progress · Blocked) / Key Decisions / Next Steps / Critical Context.

postcompact-rehydrate.mjs re-injects filesystem state; additionally the build harness could emit a resume brief in this shape from state.json — useful post-compaction and for --resume in a fresh session.

5. Issue-numbered regression tests

Their convention: regressions live in test/suite/regressions/ named <issue-number>-<short-slug>.test.ts. ArcKit's CLAUDE.md already narrates fixes by issue number (#715, #723); mirroring that as tests/regressions/ would make the drift-guard tests self-documenting.

6. Supply-chain hardening

• 7-day dependency cooldown: .npmrc min-release-age=7 + matching dependabot cooldown (with documented --min-release-age=0 override for urgent security patches). Cheap and on-brand for a UK-public-sector governance toolkit; applies to the paperclip extension's npm dependencies.
• Checksum-verified installer (their install.sh verifies SHA-256 of the versioned release).

Interop finding: prime-agent as a near-free tenth distribution format

Prime-agent implements the Agent Skills standard and discovers skills from .agents/skills/ in the project directory and ancestors — exactly where arckit init --ai codex already scaffolds ArcKit's Codex skills. An ArcKit project initialised for Codex should be usable from prime-agent today with zero converter work (it also reads ~/.claude/skills and ~/.codex/skills via its skills settings array). Needs a quick manual test before claiming support.

Suggested starting order

1 (build gates) and 2 (incomplete-build nudge) first; 3 (rollback snapshots) after auditing what autoresearch already records.