research: draft goal-centred replan ablation

[loopx-agent]

Goal And Delivered Outcome

The original Goal already appears in replan context, but presence alone does not establish that a long-running agent reassesses its most consequential remaining gap. This bilingual Draft proposes a late-checkpoint study that separates whole-goal reflection (G) from a source-bound reading index (C), while retaining the original Goal and complete ordinary input in every cell.

The proposal now specifies the actual delivery boundary: a current, due-replan CLI response may occur inside preparation while checkpoint capture is correctly gated. The advisory must preserve that current response and cannot use source-checkpoint commands as destination admission. Intended base: main. This increment changes documentation only; no shipped runtime, configuration, scoring or experiment behavior changes.

Author Declaration

• Written by: OpenAI Codex, model agent, using the LoopX Agent contribution identity.
• Specification: the maintainer-requested research question above; existing measurement/episode acceptance in docs/architecture/rfcs/long-horizon-harness-benchmark-research-program-v0.md.
• Ownership: benchmark-toolkit research with the native provider; reuse existing TypeScript replan/context and checkpoint selection owners. G/C labels remain local to the study.

Criterion	Disposition	Evidence / limit
Concrete reflection/context candidates and a four-cell design	implemented in proposal	Both study documents retain the original Goal and distinguish guidance from machine obligations.
Late-point selection, native outcomes, evidence adoption and full costs	specified; empirical work open	Admission, capacity, common-window and measurement sections.
Current-response framing and failure/replay contract	engineering prototype qualified	Disposable isolated File/SQLite CLI checks preserve every original member, envelope and identity; baseline/off parity and same-guard retry verified.
Actual model receiver and warm-checkpoint journey	partial	Older-source synthetic model journeys characterize client-visible delivery and settled follow-through. Current-source model qualification did not complete because its client transport failed before tool use; receiver/client versions remain a confound.
Scientific benefit or a product default	unqualified	Requires prospectively frozen late-point comparisons and independent assessment.

Scope And Remaining Author Work

The disposable prototype invokes the real destination CLI once and returns one valid JSON frame. It binds the original response and separate advisory, preserves existing projection freshness/omission facts, and falls back to the committed raw response on composition failure without replaying an effect. It is not installed in a runner or offered as a public API. No fixtures, private outputs or model transcripts are published.

Current CLI qualification must cover the whole checkout and actual interpreter; a few unchanged owner files are insufficient. A generated file named model-visible is also insufficient: require the actual client-visible tool response, accepted replan/writeback and subsequent useful work. Retain failed startup/recovery costs. Synthetic transport evidence is distinct from native checkpoint recovery and benchmark outcome evidence.

Remaining author work: qualify the actual pinned warm receiver/model-client path, freeze late-point manifests and treatment bytes before outcomes, then run the authorized matched continuation study. Common fresh reconstruction already includes the original objective, failed alternatives and retained best; G estimates additional reflection beyond that baseline. Do not borrow an existing fresh arm as an unplanned control or exceed the available worker slots and original cumulative budget.

The first C index only organizes already-delivered facts; absent stride-only follow-through and candidate/incumbent bindings remain unknown. No extra retrieval, summary model, stride activation or parallel decision owner is introduced. Keep Draft while this concrete author work is unfinished. Review, merge, installation and scientific acceptance remain separate.

Future-facing pass: reuse the typed readers and pre-decision selector; defer a production abstraction until a real receiver callsite is qualified. Any later caller-facing option must include its necessary frontend/CLI/Lark configuration, readback and disable journey through the existing owner.

Validation

Documentation head: b48a2c6450bbaeb3ac511cf7d42500a38870ae07, merged from main 105c4bc15091afac832495167cb98a1389951a04. Current isolated CLI checks use the pinned receiver checkout, separate from the older-source model characterization. No benchmark, native registration, grading or scientific G/C cell was launched for this increment.

Check kind	Result	Scope / limitation
static	passed	Relative links/anchors, Markdown fences, four study cells, bilingual meaning, public/private boundary and whitespace checks. Only the two study documents and an existing lower-page benchmark-guide link differ from main.
static	passed	Development-time semantic advisory: no source vocabulary change. Bilingual RFC staged/head checks pass; no existing RFC pair changed. Branch commits carry DCO sign-offs.
integration	passed, bounded	Current isolated File/SQLite CLI framing, original-value preservation, identical C across G, off/ordinary/error parity, destination mismatch, composition failure and lost-display same-guard recovery. No additional debit or effect replay.
real_entrypoint	partial	Older-source synthetic model preparation, inner replan, settlement and subsequent artifact observed. An ambiguous synthetic metric denominator remains unqualified. Current-source model transport failed before any model tool call; this is not a current receiver or scientific pass.
empirical	not run	Real late-checkpoint G/C continuation, native outcome, full marginal study costs and scientific utility remain open.

Frontend / Visual Evidence

UI impact: none. Plain Markdown study documents and a lower-page adapter-guide link. The current update leaves the opening viewport unchanged. No packaged frontend, installed setting or browser journey is claimed.

Type of Change

• Documentation update

LoopX Area

• Benchmark boundary (research protocol only)
• Public docs

Technical Direction

Existing overall roadmap S11 and long-horizon measurement/episode acceptance. No new roadmap, shared protocol vocabulary, runtime option or second Goal/context owner.

Shared-authority RFC fixture impact

N/A. No authority store, migration, provider or runtime-routing change.

Boundary Checklist

• Public-safe diff/body; no raw evidence, credentials, private identities, internal links or local paths.
• Proposal and disposable engineering observations remain separate from installed behavior and scientific acceptance.
• No existing English/Chinese RFC pair changed; study documents are bilingual.
• UI impact is none; this update does not alter the opening viewport.
• All branch commits include DCO sign-offs.
• Actual warm receiver and empirical work remain concrete unfinished Draft author work.