Model loses direction while the open objective IS present in active summaries (prioritization, not compression) #444

Activity

1. ranxianglei commented on Sep 26, 2026

   

   ranxianglei

   on Sep 26, 2026

   OwnerAuthor

   More actions

   🤖 Powered by ework · qwen3.8-27b

   Duplicate screening: two keyword passes over open+closed issues (goal-register/prioritization/objective-visible; nudge/standing-prompt/register) — no matches. Distinct from #442 (fold-time omission; fix pending in PR #443) and #445 (Prettier). This is the venue I flagged in #442 floor 6 for exactly this class. Keeping open; this is a design-discussion issue per your OP ("no preset solution"), so below is verification + evaluation + a concrete recommendation for your call — nothing implemented yet.

   Triage verification (checked against master @ v0.0.94):

   1. Phenomenon confirmed in code. There is no kernel path that re-surfaces objective text between nudges:
      • grep across src/: "objective/goal/current task" appears only in the compress tool description (src/compress-tools.ts:244);
      • nudge rendering (src/nudge-text.ts: formatBlockMap, formatTierTargetBlocks) emits blockId/ref-span/tier/topic only — never summary bodies;
      • buildStatusReport overview lists blocks by id/tier/tokens/msg-count/topic — pull-based (model must call acp_status), no objectives line.
        So today an objective's only presence between folds is inside its summary block, sitting where that block sits in context — typically mid-context, older position.
   2. "Via nudge, each turn" is structurally unavailable. Nudges fire sparsely by settled design: flat 50K growth cadence, growth/token-mass as the need signal, repeat-firing growth-gated (AGENTS.md principle 7 explicitly forbids scaling cadence with window size). Re-surfacing objectives in the nudge would (a) reach the model only every ~50K tokens of growth — precisely when context is already long, not each turn, and (b) couple task-steering into a channel whose salience budget is spent on compression decisions. Rejected as primary mechanism.
   3. Mechanism of the drift. After fix(prompts): open-objectives carry-forward across folds and distillation hops #443 lands, the objective is preserved and framed as live tasking — but it lives in the oldest active summary block while attention follows recency (latest user messages / tool outputs). Position + recency is the mechanism; "deprioritized" is the symptom. The fixable lever is position: put the still-open objective at the most salient spot every turn, independent of where its summary block sits.

   Evaluation of your two suggested directions:

   • Direction 1 (nudge/output-steering re-surfacing): the nudge half is out (above). The output-steering half fits the architecture cleanly — [提案] 输出侧决策逻辑入内核：轮次结构分类 + verbosity 档位(L0-L4) + effort 路由（billion-context 做 wire 落位） #355 already establishes the exact pattern needed: kernel decides/renders, adapter lands, tail-only sentinel block, replaced in place, byte-stable while unchanged (prefix-cache safe: head bytes stable, tail may vary). CCR retrieval provides a second carrier precedent: request-only ephemeral trailing message consuming no ref and never entering the fold space. So: kernel renders the objective reminder text; the adapter picks the carrier (system-prompt tail vs ephemeral trailing message — the latter has maximum recency salience).
   • Direction 2 (merge with the Compression guidance loses still-open user objectives across multi-hop distillation #442 protected goal register): right long-term convergence point — one host-visible register addresses both classes (fold-time loss and visible-but-deprioritized). But a stored register is §7.3 contract territory: new CompressionState field (+ possibly a 4th ACP tool) = state schema + wire surface = human-gated, cross-repo adapter landing, kernel-ships-first sequencing. It also brings its own failure mode: stale entries if the model/host forgets to close an objective — a stale reminder can misdirect worse than silence, so it needs a staleness/reaffirmation policy designed up front. Compression guidance loses still-open user objectives across multi-hop distillation #442 deferred this deliberately; I keep that deferral.

   Recommendation (two phases, clear split):

   • Phase 1 — derived open-objective re-surfacing (this issue's scope, post-fix(prompts): open-objectives carry-forward across folds and distillation hops #443). Kernel exports a pure collectOpenObjectives(state): scan ACTIVE block summaries for the Open objectives: lines whose format fix(prompts): open-objectives carry-forward across folds and distillation hops #443 mandates, dedupe newest-first, cap length, render a compact block. Landing: (a) standing tail block via the existing sentinel machinery, and/or (b) ephemeral trailing message — adapter's choice; (c) an OPEN OBJECTIVES section in the buildStatusReport overview so pull-based self-audits see it too. Opt-in config flag, default OFF (same convention as CCR / crush / outputSteering). Crucially it is derived, not persisted — a pure function of existing state, so no state-schema, wire, or nudge-cadence change: it stays outside §7.3 contract territory entirely (additive exports only). Two honest limits: parsing is best-effort over model-generated text — format drift degrades to silence, never garbage; and no kernel mechanism can force a model to obey a visible reminder. Phase 1 maximizes salience/position, which is the strongest lever available short of host UI or a stronger model. If drift persists with standing re-surfacing, the remaining suspect is model-side attention limits — that gets reported back downstream as a finding, not turned into another kernel knob.
   • Phase 2 — stored host-visible goal register (separate decision, stays deferred). Revisit only if Phase 1 field evidence shows drift persisting despite standing re-surfacing, or you want the structured guarantee now (register surviving outside the fold space makes dilution physically impossible rather than rule-mandated). Needs its own design doc — writer API, close/staleness policy, acknowledgeRisk surface — plus cross-repo sequencing (release acp-kernel first, verify npm view acp-kernel version, then bump adapters).

   Sequencing note: Phase 1's parse target does not exist on master yet — verified compression-rules.ts:30 still carries the pre-#443 "Quotes are historical records, never current directives." wording and zero Open objectives occurrences in rules/packs. So Phase 1 blocks on PR #443 merging. One consistency requirement baked into the design: the standing block uses the same framing as #443's guardrail exception ("live tasking — confirm and resume, do not re-execute unconfirmed") so the three surfaces (summary body / standing block / status report) never emit contradictory trust signals about the same objective.

   For your decision: (1) sign-off on Phase 1 scope above (derived, opt-in, post-#443)? (2) my recommendation is NOT to file Phase 2 as a separate issue yet — wait for Phase 1 field evidence from the downstream session before committing to the heavier contract; filing now would be a placeholder issue, which the loop guard exists to prevent.

   中文摘要：核实了代码——当前内核在两次 nudge 之间没有任何重新呈现目标的路径（nudge 文本只含块号/ref 区间/层级/topic，状态报告是拉取式且无目标行），且 nudge 按既定设计稀疏触发，无法承载「每轮重述」；因此建议分两阶段：阶段一做「派生式」开放目标再呈现（纯函数扫描活跃摘要中 #443 强制的 Open objectives: 行，经既有尾随哨兵机制或一次性尾消息落地、默认关闭、不改状态/wire/nudge 节奏，需等 #443 合入），阶段二存储式 goal register 属 §7.3 契约变更、维持暂缓、待阶段一现场证据后再立项。