Independent use will determine whether Arc’s persistent project memory, isolated worktrees, planning, and Claude–Codex handoffs materially reduce coding-agent degradation across long sessions and context compactions.

state: expiredheat: lowuncertainty: highknownscott: lowagent-memory coding-agents agent-harnesses long-running-orchestrationZie619

What is this?

Arc is a personal project shared by Zie619 after more than a year of daily Claude Code use; its earliest public artifact describes it as “a terminal orchestrator for AI coding agents.” The case attributes persistent project memory, isolated worktrees, planning, and Claude–Codex handoffs to Arc, targeting the documented problem of context degradation across long sessions and compactions. The supplied results establish that context rot and cross-session memory are active coding-agent problems, but provide no independent evaluation of Arc itself, so its material benefit remains unverified.

Why it matters to Scott

Scott already holds Arc’s core architectural position in `long-running-agents` and `context-engineering`: durable external state, checkpoints, bounded workers, and context hygiene should carry work across session death and compaction. The radar also tracks close variants on the Rune and BDFL pages; without independent results, Arc is another unverified implementation rather than evidence that extends or changes Scott’s position.

ip:framework.long-running-agentsip:framework.context-engineeringip:concept.checkpoint-disciplinedev:concept.agent-authored-context-compactiondev:concept.resumable-agent-job-control-planeradar:rune-persistent-coding-contextradar:bdfl-supervisor-workflow-validationradar:concept.agent-memoryradar:concept.long-running-orchestration

queries asked of Scott's wikis

persistent memory across agent compactions
coding-agent context degradation and recovery
isolated worktrees for parallel coding agents
planning harnesses for long-running agent tasks
Claude and Codex model handoff orchestration
independent evaluation of coding-agent harnesses

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (7) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	I've been using Claude Code daily for over a year. This is the personal project I now write all my code with — decided to share it. ClaudeAI	eliadkid	0	2
🟧 echo.github ⭐	Earliest public artifact found: the initial Arc release commit. Its message describes “a terminal orchestrator for AI coding agents” where “	Zie619	—	—
🟠 reddit	Compaction keeps what we built and forgets what we rejected ClaudeAI	Prestigious_Grape786	0	6
🟠 reddit	How do you handle it when a decision you made early in a project turns out wrong eight sessions later? ClaudeAI	SSShken	0	20
🟠 reddit	Four months of building a city map out of Telegram channels ClaudeAI	Winter_Hornet704	0	1
🟠 reddit	Claude Code hook that puts the recorded reason in front of the agent before it edits ClaudeAI	Illustrious-Return-4	1	2
🟠 reddit	split history split context ClaudeAI	Miki800	1	2

Interpretation history

2026-09-06T07:22:07Z
Repeated reviews have yielded adjacent failure anecdotes and memory patterns, but no independent Arc use or comparative outcomes; the latest refresh adds no substantive evidence. This episode has faded without a concrete pending validation milestone, so retire active monitoring rather than interpret silence as disproof.
2026-09-04T06:23:41Z
The latest activity is repetitive discussion of an already-known session-state failure, not independent Arc use or comparative results. After repeated reobservations, Arc remains an unvalidated implementation with no evidence that its workflow improves compaction resilience.
2026-09-02T05:30:57Z
The new report adds weak negative field evidence that compaction and session forking can separate history from summarized context, reinforcing the failure Arc targets. It still neither evaluates Arc nor isolates the cause, so Arc’s claimed long-session benefit remains unvalidated.
2026-09-02T05:21:56Z
evidence attached: reddit.post.1w501y8 — The report provides negative field evidence that session forking and compaction can fragment coding-agent history, context, and memory.
2026-08-31T12:33:27Z
The latest reobservation is only minor, unfavorable engagement churn and adds no independent Arc use, evaluation, or outcomes. The adjacent memory implementations still define useful test criteria, but Arc’s claimed resistance to long-session degradation remains wholly unvalidated.
2026-08-29T11:33:44Z
Whence adds a concrete adjacent implementation that injects stored decision rationale before edits, strengthening the evaluation pattern around external memory. It neither tests Arc nor provides outcomes for compaction recovery, stale-memory revision, or long-session performance, so Arc’s material benefit remains unverified.
2026-08-29T11:23:23Z
evidence attached: reddit.post.1w1j0n7 — This first-party Claude Code hook is a concrete artifact for preserving decision context before edits, relevant to coding-agent memory and long-session degradation.
2026-08-28T00:27:22Z
Refreshed comments repeat established advice about external notes, bounded sessions, documentation hygiene, and code as source of truth. They provide no independent Arc use or outcome evidence, leaving the implementation claim untested despite continued discussion of the underlying problem.
2026-08-26T00:25:30Z
An independent real-project anecdote lends weak support to Arc-like planner/implementer separation and verification as ways to reduce repeated mistakes. It does not evaluate Arc, persistent memory, compaction recovery, or measured long-session outcomes, so the core hypothesis remains untested.
2026-08-26T00:23:12Z
evidence attached: reddit.post.1vyg4vf — The evidence-first planner/implementer split directly contextualizes whether structured roles and verified task handoffs reduce coding-agent degradation.
2026-08-25T23:40:18Z
The refreshed discussion reinforces that durable memory needs explicit amendments, supersession, and a clear source-of-truth policy to avoid preserving stale decisions. It adds evaluation criteria but still provides no independent use or outcome evidence for Arc.
2026-08-25T17:41:58Z
The new anecdote adds an important caveat: durable project memory can preserve obsolete decisions and actively mislead agents unless corrections invalidate or supersede prior state. Arc still lacks independent use or outcome evidence, so evaluation should test memory revision and provenance—not merely persistence across compaction.
2026-08-25T17:24:57Z
evidence attached: reddit.post.1vy5xfu — This provides concrete user evidence that stale project decisions and undocumented corrections degrade coding-agent performance over long sessions.
2026-08-24T10:25:14Z
Independent testimony sharpens the underlying failure mode: compaction can preserve built state while losing rejected approaches and rationale. It still does not test Arc or show that Arc’s memory and handoff workflow materially mitigates that failure, so the case remains an unverified implementation.
2026-08-24T10:22:40Z
evidence attached: reddit.post.1vwy3hw — Direct user evidence supports the open case by identifying compaction's loss of rejected approaches as a concrete long-session coding-agent failure.
2026-08-23T12:34:56Z
No independent adoption, evaluation, or implementation evidence has appeared; the reobservation is only a negligible engagement change. Arc remains an unverified instance of an architecture Scott already knows rather than new evidence about compaction-resistant coding workflows.
2026-08-23T12:28:28Z
grounded: known/low — Scott already holds Arc’s core architectural position in `long-running-agents` and `context-engineering`: durable external state, checkpoints, bounded workers,
2026-08-23T12:26:16Z
origin walked (codex/luna, conf 0.97): anchor reddit.post.1vw59w2 -> echo.github.fae746131e by Zie619
2026-08-23T12:24:58Z
case created — Arc is a concrete released workflow addressing long-session context degradation, but currently has only one low-engagement first-party report.

Decision trace

09-06 17:22expireRepeated reviews have yielded adjacent failure anecdotes and memory patterns, but no independent Arc use or comparative outcomes; the latest refresh adds no substantive evidence. This episode has fade
09-06 17:22alert_silentThere is no new consequential delta or time-sensitive decision for Scott. Independent Arc usage or measured compaction-recovery results would justify reopening the case; discussion churn does not warr
09-06 17:22alert_routeThere is no new consequential delta or time-sensitive decision for Scott. Independent Arc usage or measured compaction-recovery results would justify reopening the case; discussion churn does not warr
09-04 16:23repriceThe latest activity is repetitive discussion of an already-known session-state failure, not independent Arc use or comparative results. After repeated reobservations, Arc remains an unvalidated implem
09-04 16:23alert_silentTwo additional comments on adjacent failure evidence do not establish a new event, evaluate Arc, or alter Scott’s implementation choices; routine review can wait for an independent user report or meas
09-04 16:23alert_routeTwo additional comments on adjacent failure evidence do not establish a new event, evaluate Arc, or alter Scott’s implementation choices; routine review can wait for an independent user report or meas
09-02 15:30repriceThe new report adds weak negative field evidence that compaction and session forking can separate history from summarized context, reinforcing the failure Arc targets. It still neither evaluates Arc n
09-02 15:30alert_silentA single ambiguous user report reinforces an already-known session-state failure without testing Arc or changing Scott’s implementation decisions; independent Arc usage or comparative long-session res
09-02 15:30alert_routeA single ambiguous user report reinforces an already-known session-state failure without testing Arc or changing Scott’s implementation decisions; independent Arc usage or comparative long-session res
09-02 15:22alert_silentA single low-engagement user report describes context/history splitting after compaction, a model switch, and a required client update, but does not isolate the cause or provide independent evidence a
09-02 15:22alert_routeA single low-engagement user report describes context/history splitting after compaction, a model switch, and a required client update, but does not isolate the cause or provide independent evidence a
09-02 15:21attachThe report provides negative field evidence that session forking and compaction can fragment coding-agent history, context, and memory.
09-02 15:21propose_attachThe report provides negative field evidence that session forking and compaction can fragment coding-agent history, context, and memory.
08-31 22:33repriceThe latest reobservation is only minor, unfavorable engagement churn and adds no independent Arc use, evaluation, or outcomes. The adjacent memory implementations still define useful test criteria, bu
08-31 22:33alert_silentNo consequential new delta occurred; the added comments and engagement changes do not alter the evidence or Scott’s decisions, so this can wait for independent usage or comparative results.
08-31 22:33alert_routeNo consequential new delta occurred; the added comments and engagement changes do not alter the evidence or Scott’s decisions, so this can wait for independent usage or comparative results.
08-29 21:33repriceWhence adds a concrete adjacent implementation that injects stored decision rationale before edits, strengthening the evaluation pattern around external memory. It neither tests Arc nor provides outco
08-29 21:33alert_silentThis is a small first-party implementation of an already-known pattern, without independent usage or comparative results; it can wait for routine briefing unless evidence appears that Arc or Whence me
08-29 21:33alert_routeThis is a small first-party implementation of an already-known pattern, without independent usage or comparative results; it can wait for routine briefing unless evidence appears that Arc or Whence me
08-29 21:24alert_silentWhence is a concrete but small implementation of an already-held pattern: persist decision rationale outside the model and inject it at edit time. The self-reported release provides no independent res
08-29 21:24alert_routeWhence is a concrete but small implementation of an already-held pattern: persist decision rationale outside the model and inject it at edit time. The self-reported release provides no independent res
08-29 21:23attachThis first-party Claude Code hook is a concrete artifact for preserving decision context before edits, relevant to coding-agent memory and long-session degradation.
08-29 21:22propose_attachThis first-party Claude Code hook is a concrete artifact for preserving decision context before edits, relevant to coding-agent memory and long-session degradation.
08-28 10:27repriceRefreshed comments repeat established advice about external notes, bounded sessions, documentation hygiene, and code as source of truth. They provide no independent Arc use or outcome evidence, leavin
08-28 10:27alert_silentThe delta is repetitive discussion rather than independent adoption, evaluation, or a material product change, so it does not merit attention before the next briefing.
08-28 10:27alert_routeThe delta is repetitive discussion rather than independent adoption, evaluation, or a material product change, so it does not merit attention before the next briefing.
08-26 10:25repriceAn independent real-project anecdote lends weak support to Arc-like planner/implementer separation and verification as ways to reduce repeated mistakes. It does not evaluate Arc, persistent memory, co
08-26 10:25alert_silentThe new evidence is an analogous single-user workflow report rather than independent Arc adoption or an outcome comparison; it refines evaluation criteria but does not change Scott’s established view
08-26 10:25alert_routeThe new evidence is an analogous single-user workflow report rather than independent Arc adoption or an outcome comparison; it refines evaluation criteria but does not change Scott’s established view
08-26 10:23alert_silentA single developer reports that a planner/diagnoser and isolated implementer/reviewer workflow reduced repeated mistakes on a four-month real project, providing anecdotal support for Arc-like separati
08-26 10:23surface_candidateA single developer reports that a planner/diagnoser and isolated implementer/reviewer workflow reduced repeated mistakes on a four-month real project, providing anecdotal support for Arc-like separati
08-26 10:23alert_routeA single developer reports that a planner/diagnoser and isolated implementer/reviewer workflow reduced repeated mistakes on a four-month real project, providing anecdotal support for Arc-like separati
08-26 10:23attachThe evidence-first planner/implementer split directly contextualizes whether structured roles and verified task handoffs reduce coding-agent degradation.
08-26 10:22propose_attachThe evidence-first planner/implementer split directly contextualizes whether structured roles and verified task handoffs reduce coding-agent degradation.
08-26 09:40repriceThe refreshed discussion reinforces that durable memory needs explicit amendments, supersession, and a clear source-of-truth policy to avoid preserving stale decisions. It adds evaluation criteria but
08-26 09:40alert_silentThe comments elaborate an already-absorbed memory-maintenance caveat and neither validate Arc nor change Scott’s established context-engineering position; routine review is sufficient.
08-26 09:40alert_routeThe comments elaborate an already-absorbed memory-maintenance caveat and neither validate Arc nor change Scott’s established context-engineering position; routine review is sufficient.
08-26 04:22sensor_dirtycomment_update
08-26 03:41repriceThe new anecdote adds an important caveat: durable project memory can preserve obsolete decisions and actively mislead agents unless corrections invalidate or supersede prior state. Arc still lacks in
08-26 03:41alert_silentThis sharpens the test criteria for persistent agent memory but neither evaluates Arc nor changes Scott’s established context-engineering position; it can wait for routine review.