Skip to content

Latest commit

 

History

History
85 lines (51 loc) · 11.9 KB

MEMORY.md

File metadata and controls

85 lines (51 loc) · 11.9 KB

Project Collaboration Memory

Only collaboration preferences, project status and reference pointers belong here. Technical contracts live in AGENTS.md/CLAUDE.md; rationale and measured evidence live in docs/status/DECISIONS.md and the handoff.

Index

Type Status Entry
user verified-active Long-running autonomous research, with P7 as the current reference domain
feedback verified-active LLM owns reasoning; infrastructure supports rather than prescribes cognition
feedback verified-active Preserve approved Prime direction and independent application ownership
feedback verified-active Research-scale source review and focused boundary checks
feedback verified-active Reconcile actual evidence before attributing errors or changing scope
feedback verified-active Inspect macro research organization and distinguish observations from progress
feedback verified-active Preserve private success/failure experience and visible native judgments
feedback verified-active Integrate and push completed work promptly; preserve unrelated changes
feedback verified-active Complete handoff directly, with clean Git and recoverable state
feedback verified-active Report throughput, costs and budget boundaries separately
feedback verified-active Fix bounded resource configuration at application responsibility
project verified-active Prime/P7 milestone, README and targeted gate witness are closed
reference verified-active Repository handoff is the authority for next actions and evidence scope

Verified Active

user — research purpose

Prime targets long-running autonomous research with understandable, falsifiable and replayable records. P7 is the current experimental domain; transfer needs its own evidence.

feedback — model reasoning and cognition

Let the LLM organize hypotheses, representations, experiments, programs, NOTES and strategy. Supply strong infrastructure, tools, actual observations and enforceable boundaries. Do not replace useful reasoning with execution-log accumulation, mandatory tiny cognition steps or hand-written solving hints. Actor programs/automatic continuation count as autonomous work; disclose outer solving intervention.

feedback — approved architecture

Do not reopen agreed Prime/native direction from old branch notes. Independent applications consume a reusable Prime library; central registration or a CLI must not become prerequisites. User model preferences and application ownership are authoritative in AGENTS and D-2026-10-10-03. Keep P2 capability work separate from migration.

feedback — proportionate verification

Review changed implementation, necessary boundary assertions and small relevant regressions. Avoid repeated release-scale/full-combination gates or unrelated suites during research. A frozen gate answers one measurement, not another hardening round. Verify behavior and deltas, not grep counts, fixture resemblance or a subagent's reported completion. Freeze reviewed source before its final packaged witness. Sources: user steering2026-09-14,2026-10-11/12.

feedback — evidence and causal attribution

A stage label or an RPC filename is not a cause. Trace the actual producer, response and dispatch state before blaming Pi or another layer. Reconcile a diagnostic with known successful runs. Search PATH and real environment before saying a resource is absent. Capture acknowledgement, values and code-path evidence; distinguish no dispatch from unknown execution. Take contracts from producers, check affected boundaries and reconcile recurring defects. Keep narrow failures narrow.

feedback — research process review

Track whether P7's macro research organization is useful, stalled or anomalous: actual hypotheses, modeling/search, experiments, feedback and revisions. Visible pixel changes are action effects, not objective progress or safety. Judge understanding by its use in later decisions. User-defined checkpoints are completed assigned targets; six levels were not a demonstrated solving ceiling. Route variation across different code/context/budgets cannot automatically be attributed to randomness.

feedback — experience, console and sharing

Preserve privately complete successful and failed research, counterexamples and useful programs. New fresh research excludes prior experience unless explicitly selected; an instructed continuation reuses its admitted checkpoint. Replay displays native saved judgments and never uses another LLM to invent missing history. HUMAN play is independent of P7 research. Keep console layouts compact, scores legible and selections persistent. Public results/projections stay separate from private archives and credentials.

feedback — integration and autonomy

Completed work should be integrated into main, committed and pushed promptly. Use subagents for concrete bounded tasks under AGENTS; judge them by delivered work and verify integration. Preserve other workers' edits and unrelated worktrees/processes. Do not repeat authorization or provider/budget questions for an already authorized preset.

feedback — final handoff

When the user says handoff, complete final closeout directly: stop this session's leftovers, persist verified facts/current judgments/history/open boundaries, correct core state and memory pointers, commit/push and leave Git clean. Do not hand the user another draft to approve. Keep scope necessary; do not create another state hierarchy. The repo AGENTS handoff contract overrides a generic skill's draft-review convention.

feedback — resource reporting

Report restoration versus new actions, per-level versus whole-game scores, attempted versus effective evaluations and source versus deployed evidence separately. Token totals are throughput; cache fractions and labelled cost estimates are different facts. Missing or partial data stays unknown; an SDK estimate is not an invoice. Deaths are not automatically research failure, but avoidability from prior evidence and actual probe authority must be checked. A solved level need not be rerun solely for a bounded gate/summary repair when exact saved evidence and a targeted witness suffice.

feedback — application budgets

A low inherited cap is an application integration problem, not a reason to ask the user to micromanage a budget. Applications must supply suitable finite execution presets. Model risk/cost declarations alone are not operator permission; finite unknown-risk research must preserve the distinction between accepting risk and establishing safety. Technical policy: D-2026-10-13-03.

project — milestone closed

The user accepted and tagged the Prime/P7 milestone. Consolidation, scoped repairs, checkpoint completion, targeted witness and bilingual README are complete. Do not restart a stopped trial, open another official card or start a new paid run from stale permissions. Current immutable tag identities, source snapshots and receipts are in docs/status/RESUME-NEXT-SESSION.md.

reference — restore without chat history

Start with project-state resume: INDEX → CURRENT-STATE → final RESUME → recent JOURNAL, then check Git and actual recorded processes. Technical decisions and old qualifications are in DECISIONS and indexed reports. Historical examples remain in the pre-closeout memory snapshot; they must not override current project status.

Current Judgment

The Prime/P7 milestone remains closed. The 2026-10-20 completion worklist is done. The peer-foundation worklist is done: P-PI-1, P-NATIVE-1, and P-CITE-1. PiRpcSession keeps PiChildProcess, and both the session and the one-shot facade reap through it. Provider native keeps native.packages@1.0.0 first and also exposes native.dci-research@1.0.0 and native.controlled-code@1.0.0. prepare_native_application is the Native door. prepare_prime_application stays the Prime door. The shipped retrieval preset can complete; exit 0 is not a judged answer. Installed provider dci-agent-lite is removed. native.dci-complete is specified and not built. The product CLI is asterion-dci. The Judge factory key still uses the name dci-agent-lite. The grid-regulation application cites a dotted clause across ordered PDF and DOCX spans and does not start Pi. Its operator desk is the single file page.html. The page asks for one regulation file, such as the 2023 stability regulation, and the fault or matter. POST /api/measures calls answer_fault. page.main passes LiveGridResearch. serve with no research object still returns 422 and does not start Pi. A supplied research object returns its reply and the passages a read opened. The option builder sets tools read,bash and max_turns "32". There is no clock cancel and no 60-second cancel. 阅读没有完成 N 秒 is elapsed time. A read of the projection's own passages.txt or contents.txt counts when the path is absolute; a path outside that directory still fails [c1525efd]. The operator interrupts. Tests do not call run. make grid-regulation serves 127.0.0.1 and starts Pi only when the form is submitted. While a search runs, the button counts seconds. stderr prints the turn, then the sentence the model wrote before the tool, then the file and line window being read, or the projection file and the search text. The question, the command, the path, the tool result, and the final answer stay off that log [b6f29c24, ab856b5c]. The reply renders as Chinese Markdown. Evidence is one paragraph or one table. The handling column is heavier type. Bold text is weight only, with no yellow fill. The reading face is PingFang SC, with installed gothic fallbacks and no remote font. A desk process started before 38029c7e still omits span text until it is restarted. The task text requires the first paragraph to answer the question. Evidence under a styled heading is one section card titled with that outline path. Clause numbers are citation labels. It does not ask for a clause number first. When a DOCX contents list at levels 1–3 matches heading styles 1–3, the clause number is placed on the heading and the contents entries are not cited. A span that is only a dotted number does not start a clause. A later model-written grid answer uses native-dci through prepare_native_application. prepare_prime_application stays the Prime door. Tests and serve without a research object do not start Pi. page.main would start Pi on submit. The third attempt of one authorized question on the 2023 stability DOCX rendered a 1917-character reply from 18 reads and 835 anchors. That reply is not an applicability judgment. Another model call is not authorized. The grid command passes --thinking off. make grid-regulation serves the page on localhost. Regulation originals under docs/ stay untracked. No package is pending. The fixed P1 preset exits 0 only when status is completed. Recorded samples include oracle.pass on p1-f8f0aa6dfe5cb466d2c64e79 and recovery-required on p1-e822af327220e58ba4c6c43d and p1-cf6c9ddd449a9710bd968e21. Prime stays a product-independent harness. DCI stays the retrieval package. A dataset benchmark, the complete example, the provider conversion, an official card, an L9 rerun, a new P application, another live retrieval, and a push of this handoff stay unauthorized until the user asks. Details are in docs/status/RESUME-NEXT-SESSION.md.

Historical Corrections

  • The old “architecture branch awaits main integration” statement is superseded: consolidation and remote archives are complete.
  • Earlier claims that Pi/configuration/resources were absent, or that a summary cap alone explained a failure, were revised after measurements. Keep the evidence-first lesson; do not restore those diagnoses.
  • Old remaining-game scheduling, partial-card and “next live test” permissions are historical. The current handoff defines what is closed and what would require a new target.