Uploaded materials get immediate fast-channel feedback and one background task card. The r0.4 path asks the material worker to publish one sourced fact first, searches only the current task's exact attachments when needed, and shows a deliverable choice only after a completed result has real suggestions. Incomplete tool streams fail safely. The follow-up 0637547 correction persists returned Usage on a failed child run, including a missing-total terminal after earlier turns.
Evidence
Mechanism: Backend 0637547 passed run 3079 (vet-and-test, trusted home-runner, all steps). The independent accounting review references read-only mixed-turn and SQLite replay; local red/green receipts remain in the task card.
Real-model quality: The single authorized r0.4 A01 at fixed Dev d38dbdc / Backend 37e2cb7 used two paid run IDs and 29,062 returned Usage tokens. The child tool stream was rejected before any tool or checkpoint was saved; no final organized answer exists. The fast response also asserted an unsupported dependency-to-phase relationship. Quality FAIL. The original SQLite/result underreported the failed child's 7,148 tokens; the later code fix does not rewrite that sample or establish model quality on the new head.
Merge Danger
Door: two-way
Blast Radius: material workflow and Usage accounting
## Summary
Uploaded materials get immediate fast-channel feedback and one background task card. The r0.4 path asks the material worker to publish one sourced fact first, searches only the current task's exact attachments when needed, and shows a deliverable choice only after a completed result has real suggestions. Incomplete tool streams fail safely. The follow-up `0637547` correction persists returned Usage on a failed child run, including a missing-total terminal after earlier turns.
## Evidence
- **Mechanism:** Backend `0637547` passed [run 3079](https://git-home.u1680832.nyat.app:30319/weichen/Agent-System-Backend/actions/runs/3079) (`vet-and-test`, trusted home-runner, all steps). The [independent accounting review](https://git-home.u1680832.nyat.app:30319/weichen/Agent-System-Dev/src/branch/feature/fast-slow-u2-u3-plan/.task/feature-fast-slow-u2-u3-plan.md) references read-only mixed-turn and SQLite replay; local red/green receipts remain in the task card.
- **Real-model quality:** The single authorized r0.4 A01 at fixed Dev `d38dbdc` / Backend `37e2cb7` used two paid run IDs and 29,062 returned Usage tokens. The child tool stream was rejected before any tool or checkpoint was saved; no final organized answer exists. The fast response also asserted an unsupported dependency-to-phase relationship. **Quality FAIL.** The original SQLite/result underreported the failed child's 7,148 tokens; the later code fix does not rewrite that sample or establish model quality on the new head.
## Merge Danger
**Door:** two-way
**Blast Radius:** material workflow and Usage accounting
Keep this PR draft. The A01 failure requires a reviewed design correction and a separately authorized new model trial; stage-7 acceptance and merge are not complete. See [Dev PR #512](https://git-home.u1680832.nyat.app:30319/weichen/Agent-System-Dev/pulls/512), [Documents PR #95](https://git-home.u1680832.nyat.app:30319/weichen/Agent-System-Documents/pulls/95), [Agent-Native PR #46](https://git-home.u1680832.nyat.app:30319/weichen/Agent-Native/pulls/46), and [Agent-Squad PR #53](https://git-home.u1680832.nyat.app:30319/weichen/Agent-Squad/pulls/53).
Why: compile the work projection against the reviewed Core vocabulary.
Evidence: exact local Core task commit materialized; no backend behavior changed in this dependency step.
Impact: Core gitlink only.
Refs: U2/U3 Task1
Why: equal-length edits and parser changes must not reuse prior results.
Evidence: content identity and provided-source tests RED to GREEN; backend suite recorded through evidence.
Impact: trusted optional document origins, content keys and source catalog; no new routing or DB schema.
Refs: U2/U3 Task2
Why: Background work needs consistent cards, events, and user reading state.
Evidence: SELF make evidence records SQLite rollback, projection, pin, and regression checks; PostgreSQL replay is pending the composition gate.
Impact: Adds an optional atomic task path and a consistent card replay cursor without schema changes.
Refs: Agent-System-Dev fast-slow-u2-u3 plan Task 3
Why: Real source-backed discoveries should reach the same task card before the slow channel completes.
Evidence: SELF receipts record runtime and manager checkpoint red-green tests, scoped source/version checks, privacy filtering, quota corrections, and SQLite/PG regressions; fixed composition and live comparison follow.
Impact: Adds a local material-only checkpoint sink with no summary model calls, uses version witnesses for source commits, and preserves existing reading and notification paths.
Refs: fast-slow-u2-u3 Tasks 6 and 7; test-only live HTTP harness bounds shared provider requests without approving HITL
Why: The live harness must use the actual client constructor and preserve MiniMax reasoning separation.
Evidence: SELF live-tag compilation passes without provider calls.
Impact: Test harness only; no production route changes.
Refs: fast-slow-u2-u3 Task 7
Why: Baseline and candidate must exercise their actual default prompts with identical route and resource limits.
Evidence: SELF live-tag compilation passes.
Impact: Evaluation harness only; normal automatic notice quality remains outside the controlled comparison.
Refs: fast-slow-u2-u3 Task 7
Why: Opening a processed source must not show changed current bytes as the quoted version.
Evidence: SELF TaskSourceContent regression fails before and passes after expected-hash enforcement; artifact regressions and live compilation pass.
Impact: Optional content hash narrows the existing authorized artifact read; controlled fixtures select a genuinely provided quotation.
Refs: fast-slow-u2-u3 source-version acceptance
Why: Text with short attachments previously completed without starting the intended background material task.
Evidence: SELF real HTTP/browser reproduction shows the absent second task; runtime, manager and server regressions pass with the existing optional port connected.
Impact: Preserves dispatch timing and makes deferred intake observable; comparison records the baseline gap without rerunning a sample.
Refs: fast-slow-u2-u3 A02 and reading acceptance
Why: Delayed notices, plugin name collisions and startup recovery exposed stale sources or silently created new attempts; deletion lost search index cleanup.
Evidence: Meaningful red regressions and the corrected complete Backend suite passed; production startup was exercised through an isolated subprocess.
Impact: Revalidate notices at delivery, retain plugin dispatch, restore index cleanup and complete all service wiring before recovery and public intake. Scope fixture processes and wait for owned transcript writes.
Refs: Agent-System-Dev feature-fast-slow-u2-u3-plan; independent review findings 1-3 and startup supplement
Why: The first real-model sample batched discoveries, overcalled scope differences and exposed process chatter as a terminal conclusion.
Evidence: Preserve the failed A01 sample; process-text regression failed before the fix, then complete Backend and focused runtime checks passed.
Impact: Bump recipe identity, qualify quantities and approvals, exclude Process parts from canonical results and capture independent historical evaluation snapshots.
Refs: Approved U2/U3 plan; independent model review and native SELF receipts
Why: Authentic quotes do not establish an unstated approval, and the evaluation wrapper must honor tighter production output limits.
Evidence: Approval-overstatement and 780-token-cap regressions failed first; complete Backend suite and fixture cap checks now pass.
Impact: Render material records from trusted labels and validated quotations, reserve interpretation for explicit finding kinds, bump recipe identity and cap the fixture without overriding smaller limits.
Refs: Approved R05; independent model assessment; original failed samples preserved
Why: A budget-stopped task retained a promise to generate a final result, while readable card errors need the full diagnosis in the execution record.
Evidence: Continuation-promise and missing-error-record regressions failed on the prior implementation; the full Backend suite now passes.
Impact: Terminal next steps require user action, stopped work has clear labels, and precise subagent errors persist outside the default task summary.
Refs: Independent A04 model assessment; approved R08 and R10
Why: A reconciled task must not retain the old queue or running next step after becoming interrupted.
Evidence: Owned queue recovery failed before the fix; RunMgr, actual App startup and evaluation output-cap checks now pass.
Impact: Recovery uses the stopped-task next step, and the mechanism fixture quotes the actual acceptance source text.
Refs: Approved recovery contract; independent terminal-state review
Why: The user approved one additional A01 sample after the original six-case batch failed its quality gate.
Evidence: Live fixture tests preserve the production output cap and reject paid requests when a third actual run exists.
Impact: Evaluation-only case selection and resource guard; no production behavior changes, no automatic replay.
Refs: User approval 2026-10-13; reported token target 60000
Why: The failed extra A01 had no accepted findings but claimed saved discoveries and future progress.
Evidence: The zero-findings regression failed before the fix; full RunMgr tests now pass.
Impact: Use truthful conclusion, remaining and next-step text while retaining strict source validation.
Refs: Independent extra A01 assessment; approved R08
Why: Resolve the component conflict using the tested combined Core lineage.
Evidence: Both exact Core histories are included in beafd753; no source conflict in App or Server.
Impact: Preserve TaskSnapshot and upstream managed settings/session checks; composition replay follows.
Refs: Core beafd753; Backend main 741da4a0
Why: Upstream now reserves model and mock settings for system authority.
Evidence: Complete merged Backend suite and live resource/cap guards pass.
Impact: Fixture-only startup configuration, same effective route and profile; no further paid calls.
Refs: Upstream scoped settings contract; U2/U3 evaluation
Why: Preserve the newest protected main for the combined review candidate.
Evidence: Exact incoming object fetched; affected checks will replay.
Impact: Retain both histories without relaxing source or task authority.
Refs: Dev main aa6364f
Why: Retain the newest protected Backend main alongside the approved work branch.
Evidence: Exact upstream object fetched; affected checks follow.
Impact: Preserve both histories and review the new model-tool guidance as a separate delta.
Refs: Agent-System-Dev U2/U3 delivery
Why: The bounded A01 model call produced a five-finding malformed payload and an abbreviated quote that strict validation correctly rejected.
Evidence: Full Backend suite passed; original paid failure and rejected source citations remain unchanged.
Impact: Advertise one to three findings per publication and continuous literal quotes, bumping material recipe identity to v7; Core authority and budget remain unchanged.
Refs: U2/U3 failed A01 assessment and proposed bounded follow-up
Why: The prior one-shot A01 result failed and its resource approval is exhausted.
Evidence: Follow-up authorization tests failed before the change and pass after; full Backend suite passes.
Impact: Only live evaluation harness gains a separate one-shot result and 30000-token run budget; production validation remains strict.
Refs: Agent-System-Backend#193; Agent-System-Dev#512
Why: A 250ms card poll missed a checkpoint saved about one millisecond before failure.
Evidence: The persistence regression failed before the fix, passed afterward, and independent read-only replay isolated the current conversation and running state.
Impact: Evaluation reads committed card events without changing model output or re-running the paid sample; U3 quality remains failed.
Refs: Agent-System-Backend#193; Agent-System-Dev#512
Why: An approved one-shot sample continued to a paid model call after a nonliteral quote was rejected, contrary to its stop rule.
Evidence: A zero-paid provider fixture failed before the guard and passed afterward; the original 14231-token deviation remains recorded.
Impact: Future controlled evaluations refuse subsequent provider requests after a checkpoint tool error; this does not reclassify the failed sample.
Refs: Agent-System-Backend#193; Agent-System-Dev#512
Why: A rejected checkpoint did not stop a later paid request, and a history-only fix could miss JSON errors or other active runs.
Evidence: Source and JSON rejection fixtures failed before the instance-level guard, then passed with fail-closed and unrelated-tool checks; independent replay matched both real failed databases.
Impact: Read-only event correlation blocks future provider requests after a checkpoint rejection; original sample failure and stop-rule deviation remain recorded.
Refs: Agent-System-Backend#193; Agent-System-Dev#512
Why: The old source index repeated most of a three-document input and its test fixture depended on that duplicate layout.
Evidence: Source prefix, ID scope, parser revision, table, truncation and mechanism tests failed before the change and pass after; full runmgr tests pass.
Impact: Compact the material prompt without widening quote authority, and version the recipe so old results are not reused.
Refs: Agent-System-Backend#193; Agent-System-Dev#512
Why: A source-valid analysis in the approved A01 sample promoted a requirement into an approved decision.
Evidence: The observed claim and six display fields failed before the guard, then passed with negative/question and mixed-clause regressions; full runmgr tests passed.
Impact: Keep exact material records while rejecting model-authored affirmative approval assertions before persistence; final model quality remains unverified.
Refs: Agent-System-Backend#193; Agent-System-Dev#512
Why: Material runs could spend the remaining budget on checkpoints and leave no final response.
Evidence: go test ./... -count=1; go test -race ./internal/runtime -run TestMaterialFinalReserve -count=1; offline replay of the prior A01 usage boundary.
Impact: Material subagents enter final-only mode or stop before another provider request; ordinary subagents retain their existing behavior.
Refs: U3 quality revision r0.2, Task 4
Why: Main advanced its account search and embedded Core contracts while the U3 candidate remained in review.
Evidence: go test ./... -count=1 on the merged working tree; Core, UI and Frontend merged tests passed separately.
Impact: Preserve the U3 work and account search while pinning the merged Core contract.
Refs: U3 quality revision r0.2, Dev task-sync conflict
Why: The isolated fixture passed integer settings to a float64-only reader, so the intended output cap and retry policy were ineffective.
Evidence: numeric-settings regression failed before and passed after; go test ./... -count=1 and live-tag fixture tests passed.
Impact: Mechanism and future explicitly approved model evaluations use their stated per-call output cap and disabled retries.
Refs: U3 quality revision r0.2, browser mechanism failure
Why: Independent review found false approval rejections, duplicate checkpoint counting and final-only reminder changes to evaluator grouping.
Evidence: Three focused regressions failed before fixes and passed after; go test ./... -count=1, race and live-tag fixture checks passed.
Impact: Valid questions and denials persist, only distinct checkpoint IDs count, and final-only requests keep their original run fingerprint.
Refs: U3 quality revision r0.2, independent final review
weichen
changed title from [WIP] Persist source-bound work through recovery and result delivery to [WIP] Persist sourced work and reserve a final answer2026-10-13 11:03:11 +00:00
Why: Reusing an exhausted authorization label would overwrite or misattribute the next controlled quality sample.
Evidence: New label/output and 30000-token guarded-run tests failed before implementation and passed after; full Backend Go suite passed.
Impact: The r0.2 A01 harness can be used only after separate resource approval, with an independent output name and existing two-run guard.
Refs: U3 quality implementation plan r0.2, Task 6 preparation
Why: One paid A01 run split into multiple prompt-fingerprint groups, and persisted Usage omitted real checkpoint tool calls.
Evidence: Stable-run grouping, missing-ID refusal and tool-call usage regressions failed before and passed after; full Go, race and live-fixture checks passed.
Impact: Future controlled evaluations count per-run calls and deadlines by host run identity and persist tool-call usage. The failed r0.2 A01 record is unchanged.
Refs: U3 r0.2 A01 independent replay, Task 6
Why: The current main adds private project member event handling that must coexist with U3 task projections.
Evidence: Resolved cardsvc and SSE conflicts; corrected atomic project-event proof placement after a compile RED; go test ./... -count=1 passed.
Impact: Preserve both private project visibility and task checkpoint behavior; no real-model rerun.
Refs: Backend PR #193; Dev task feature/fast-slow-u2-u3-plan
Why: The r0.2 model sample assigned interface-dependent requirements to a phase without source support.
Evidence: New source-count and set-bound tests failed before implementation; budget-item and internal-size regressions failed then passed; go test ./... -count=1 passed.
Impact: Add a pure, conservative normalizer and pin Core c722c16; no checkpoint submission or provider behavior changes yet.
Refs: Dev task feature/fast-slow-u2-u3-plan; r0.3 evidence implementation plan Task 2
Why: Valid but underdetermined model claims must become pending without losing the batch or leaking unsupported status text.
Evidence: Mixed-batch, source-input binding, approval-status, tool-receipt and prompt regressions failed before changes; go test ./... -count=1 and selected live-tag fake-provider tests passed.
Impact: Validate all task-bound sources, normalize once, return a safe receipt, and version material processing as v9; no extra provider call.
Refs: Dev task feature/fast-slow-u2-u3-plan; r0.3 evidence implementation plan Task 3
Why: The generated TypeScript mirror must keep evidence helpers private to the module.
Evidence: Backend go test ./... -count=1 passed with Core b474394; frontend knip fix verification follows after regeneration.
Impact: Nested Core gitlink only; no runtime semantic change.
Refs: Backend PR #193; Dev task feature/fast-slow-u2-u3-plan
Why: Browser-level U3 validation must render a real host-normalized needs_user finding, not only source records.
Evidence: The new Playwright assertion failed against the previous fixture; selected live-tag mechanism and stop-rule tests pass.
Impact: Test-only fake stream adds an unsupported membership claim with source-indexed premises; no provider call.
Refs: Dev task feature/fast-slow-u2-u3-plan; r0.3 plan Task 4
Why: Independent review found negated short quotes, cross-project counts and unstructured batch placement could be shown as facts.
Evidence: Each counterexample failed before repair; Backend go test ./... -count=1 and selected live-tag mechanism/stop tests passed after repair.
Impact: Require unambiguous source sentence context and matching explicit scopes, keep only cautious ordinary analysis without structured evidence, neutralize unsafe status text.
Refs: Dev task feature/fast-slow-u2-u3-plan; r0.3 review P1 findings
Why: Independent re-review found single-word decision quotes, project-prefix scope mixing, revoked totals and approval status bypasses.
Evidence: Each new counterexample failed before repair; Backend go test ./... -count=1 and selected live-tag fake-stream tests passed after.
Impact: Display complete visible source sentences, parse only bounded positive count clauses, and set material progress text from the host phase.
Refs: Dev task feature/fast-slow-u2-u3-plan; r0.3 second review Important findings
Why: Independent review showed a truncated visible prefix, revised count and same-source shortcut could still produce unsupported claims.
Evidence: Each counterexample failed before repair; positive same-named scope, Backend go test ./... -count=1 and live-tag fake-stream tests passed after.
Impact: Require a visible sentence boundary or complete block, reject unrecognized count/revision segments, and require explicit common scope for unique membership.
Refs: Dev task feature/fast-slow-u2-u3-plan; r0.3 third review Important findings
Why: The fourth independent review found extra short quotes in direct relations and stale counts revised with an unfamiliar unit.
Evidence: Both inputs failed before repair; Backend full Go and selected no-paid live-tag mechanism tests passed after.
Impact: Direct relations accept one complete source, and count proof rejects unsupported numeric segments while preserving the bounded A01 grammar.
Refs: Dev task feature/fast-slow-u2-u3-plan; r0.3 fourth review Important findings
Why: A model could put an affirmative approval before a late caution word and have the unsupported claim shown as analysis.
Evidence: Approval-behind-caution and task-prompt tests failed before repair; Backend full Go and no-paid live-tag mechanism suites passed after.
Impact: Keep a small set of cautious qualitative findings, replace all other free analysis with pending, clear model ancillary prose and bump material recipe to v10.
Refs: Dev task feature/fast-slow-u2-u3-plan; r0.3 fifth review Important finding
Why: A future separately approved model trial must not reuse the consumed r0.2 label or overwrite its raw result.
Evidence: New label/output and 30000-token guard tests failed before implementation, then selected live-tag tests and Backend full Go suite passed.
Impact: Zero-paid evaluation entry only; its name does not grant resource approval or issue provider requests.
Refs: Dev task feature/fast-slow-u2-u3-plan; r0.3 plan Task 5
weichen
changed title from [WIP] Persist sourced work and reserve a final answer to [WIP] Verify source-backed stage judgments before publication2026-10-13 18:20:36 +00:00
Why: wrapper-side tool deltas raced with serialized run events and could disappear under load.
Evidence: TestToolCallDeltaBeforeExecution race reproduced on the old path, then passed ten race runs with Agent-Native e86e74b; full Backend vet and tests pass.
Impact: one owner emits conversation events; Backend pins the Agent-Native event identity contract.
Refs: Agent-Native PR #46; internal/runtime/llm_direct.go
Why: combined filename and Chinese-keyword queries returned no hits even when the uploaded material was present, delaying useful background findings.
Evidence: combined-query, owner/project, duplicate, resource-policy, version-drift and excerpt RED→GREEN tests; SQLite and isolated PostgreSQL pass; Backend full vet and tests pass with the Agent-Native pin.
Impact: only a no-hit material subagent query may read a verified current-task attachment snapshot; general knowledge retrieval remains unchanged.
Refs: U3 r0.4 design; internal/runmgr/material_retrieve.go
Why: the background model spent its first turns searching already provided text and failed before saving any staged finding.
Evidence: first-finding and two-phase RED→GREEN tests; SQLite and isolated PostgreSQL pass; full Backend vet and tests pass; v11 prompt grew 337 UTF-8 bytes in the fixed fixture.
Impact: material recipe v11 asks for one short sourced fact before optional search without changing the checkpoint schema or adding a separate model call.
Refs: U3 r0.4 design; internal/runmgr/material.go
Why: failed or ungrounded material notices still asked nontechnical users to choose deliverables and could act on invented model options.
Evidence: failure, partial, no-suggestion, forged-marker, source-drift, reload and host-option RED→GREEN tests; race, isolated PostgreSQL, full Backend vet and tests pass.
Impact: only a completed source-bound result with 3-4 bounded suggestions can offer a choice; failed or stale notices preserve truthful status and block unsolicited delegation.
Refs: U3 r0.4 design; internal/runmgr/material_notice.go; internal/runtime/material_choice_notice.go
Why: the controlled A01 wrapper still allowed another paid request after a malformed tool terminal with returned Usage.
Evidence: local SSE RED→GREEN tests cover invalid JSON, missing DONE, late errors, missing Usage, rejected checkpoints, cross-run stop and record-write failure; live-tag fake tests and full Backend vet/tests pass.
Impact: diagnostic metadata and batch-wide fail-closed stop only; no new paid authorization or production route change.
Refs: U3 r0.4 design; internal/server/work_checkpoint_eval_test.go
Why: The Dev pin gate requires every Agent-Native consumer to use the same exact commit; Agent-Squad now pins the U3 r0.4 stream guard.
Evidence: GOWORK=off go vet ./... && GOWORK=off go test -count=1 ./... passed with this pin.
Impact: Backend dependency gitlink only; no source or public contract change.
Refs: Agent-Squad PR #53; U3 r0.4 Task 6
Why: The task-owned mechanism must reproduce stage progress, partial and empty failures, and source-bound choices without a paid provider call.
Evidence: Backend vet/full Go tests and live-tag terminal tests passed; six task-owned HTTP/Playwright cases passed with this fixture.
Impact: Evaluation fixture only; model comparison still disables notice turns unless the explicit mechanism browser flag is set.
Refs: U3 r0.4 Task 6
Why: Missing Usage must stop a paid sequence before a tool terminal is executable; exact attachment names and host-owned choices must also remain reliable.
Evidence: Each reported behavior had a failing regression before its fix. Full Backend vet/Go tests, live-tag budget tests, ten race replays, and six task-owned HTTP/Playwright scenarios passed with the aligned Agent-Native and Agent-Squad pins.
Impact: Evaluation terminal accounting becomes durable before delivery; material fallback matches bounded filenames; ready choices use host options without redundant model fields. No new model calls or budget authorization.
Refs: U3 r0.4 independent review Critical 1 and Important 3-4
weichen
changed title from [WIP] Verify source-backed stage judgments before publication to [WIP] Deliver reliable sourced material checkpoints2026-10-14 04:00:56 +00:00
Why: Backend and Agent-Squad must consume the Agent-Native revision that rejects empty tool_calls and function_call terminals.
Evidence: Full Backend vet and Go test suite passed with both exact gitlinks.
Impact: Dependency gitlinks only; no Backend source change.
Refs: Agent-Native PR #46; Agent-Squad PR #53; U3 r0.4 review closure
Why: Preserve the accepted U7 account transaction and identity safeguards while retaining U3 material checkpoint startup ordering and runtime fixes.
Evidence: The shared app.go edits are in separate startup/account-maintenance sections; full Go and parent composition checks will be rerun after this merge.
Impact: Includes mainline account changes without changing U3 public contracts or model-call budget.
Refs: U3 r0.4 resource-preparation sync; Dev main 05b997b
Why: The separately approved r0.4 trial must use its own bounded authorization label and output file, without reusing exhausted r0.3 artifacts.
Evidence: Both new label tests failed before the change and passed afterward; full Backend vet/Go suite and selected zero-paid live-tag budget/evaluation tests passed. No real provider call was made.
Impact: Evaluation-only selector and output name; production model routing, budgets and runtime behavior unchanged. Trial remains gated on fixed Dev CI and preflight.
Refs: U3 r0.4 one-time A01 resource approval in this conversation; Dev PR #512
Why: A01 returned paid Usage on a rejected tool terminal, but the failed child run persisted zero because the runtime did not emit message_end.
Evidence: RED then GREEN for failed-terminal and persisted-child regression tests; SELF receipt /tmp/fast-slow-u3-r04-20261014/usage-fix-backend-verification.json records vet, full tests, and zero-paid live-tag checks.
Impact: Reconcile the returned cumulative Usage before persisting a failed child run; preserve actual tool-call counts. The original A01 quality result remains FAIL.
Refs: U3 r0.4 A01; /tmp/fast-slow-u3-r04-20261014/A01-independent-assessment.md
Why: A failed terminal without total_tokens can follow a successful tool turn, leaving the cumulative reported total stale even though prompt and completion include both turns.
Evidence: Mixed-turn persisted child test failed before the correction and passed after; SELF receipts usage-fix-mixed-turn-red-diagnostic.json, usage-fix-mixed-turn-green.json, and usage-fix-mixed-turn-full.json.
Impact: Reconcile against the larger of cumulative total or prompt-plus-completion while preserving cache and tool counts. No real-model sample was rerun.
Refs: U3 r0.4 A01; independent review /tmp/fast-slow-u3-r04-20261014/usage-fix-independent-review.md
Why: The first material reply could stream an unsupported numeric relationship before the sourced background card had any verified finding.
Evidence: Task 2 red/green SQLite run tests cover parent provider suppression, queue/reuse, completed source binding, focus retention, squad entry, failed/unknown state, and attempt ordering; full Backend vet/Go/race SELF receipt /tmp/fast-slow-u3-r05-task2-backend-final-v2.json exited 0.
Impact: Prestarted material sends receive one server-owned status message with zero parent model calls; small-text and explicit completed follow-up paths retain their existing runner. No public contract or model budget increases.
Refs: U3 r0.5 approved design and plan; Backend PR #193
Why: The controlled A01 record retained raw prompt/tool payload while lacking enough transport classification to explain the rejected tool terminal.
Evidence: TestEvaluationRecordsContentFreeTerminal and error-bearing terminal regressions failed before the change and passed after; full Backend vet/Go and selected zero-paid live-tag SELF receipt /tmp/fast-slow-u3-r05-task3-backend-full.json exited 0.
Impact: Pin Agent-Native 6a1dcdd, retain per-call Usage and safe terminal facts, remove future raw messages/tool declarations/terminal snapshots, and clear tools on an error terminal before forwarding. Historical trial artifacts remain untouched.
Refs: U3 r0.5 approved design and plan; Backend PR #193
Why: The safe acknowledgement must be observed before fake findings arrive.
Evidence: Dev evidence receipt for ten browser scenarios and isolated PostgreSQL/Redis tests.
Impact: Changes only the task-owned zero-paid fixture delay.
Refs: U3 r0.5 implementation plan Task 4
Why: The Backend composition must use the same Agent-Native revision through Squad.
Evidence: Dev task receipt for full Backend vet and Go test with pinned dependencies.
Impact: Updates only the Agent-Squad gitlink.
Refs: U3 r0.5 implementation plan Task 5
Why: A provider-controlled finish reason could leak arbitrary content into the supposed content-free audit file.
Evidence: Red then green terminal-label regression and full Backend vet/Go test in Dev receipts.
Impact: Only evaluation records map unknown terminal labels to a finite class; tool execution rules are unchanged.
Refs: U3 r0.5 independent review finding
Why: The stage-6 browser proof must cover the real queued and failed-start user states.
Evidence: Two Playwright paths failed before fixture support, then passed; full Backend vet and Go tests passed.
Impact: Adds only task-owned mechanism fixture behavior under the live test tag.
Refs: U3 r0.5 independent review finding
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Summary
Uploaded materials get immediate fast-channel feedback and one background task card. The r0.4 path asks the material worker to publish one sourced fact first, searches only the current task's exact attachments when needed, and shows a deliverable choice only after a completed result has real suggestions. Incomplete tool streams fail safely. The follow-up
0637547correction persists returned Usage on a failed child run, including a missing-total terminal after earlier turns.Evidence
0637547passed run 3079 (vet-and-test, trusted home-runner, all steps). The independent accounting review references read-only mixed-turn and SQLite replay; local red/green receipts remain in the task card.d38dbdc/ Backend37e2cb7used two paid run IDs and 29,062 returned Usage tokens. The child tool stream was rejected before any tool or checkpoint was saved; no final organized answer exists. The fast response also asserted an unsupported dependency-to-phase relationship. Quality FAIL. The original SQLite/result underreported the failed child's 7,148 tokens; the later code fix does not rewrite that sample or establish model quality on the new head.Merge Danger
Door: two-way
Blast Radius: material workflow and Usage accounting
Keep this PR draft. The A01 failure requires a reviewed design correction and a separately authorized new model trial; stage-7 acceptance and merge are not complete. See Dev PR #512, Documents PR #95, Agent-Native PR #46, and Agent-Squad PR #53.
[WIP] Persist source-bound work through recovery and result deliveryto [WIP] Persist sourced work and reserve a final answer[WIP] Persist sourced work and reserve a final answerto [WIP] Verify source-backed stage judgments before publication[WIP] Verify source-backed stage judgments before publicationto [WIP] Deliver reliable sourced material checkpoints