Batch 23 ContextPilot exact-publication audit

Disposition

Hold. The new publication again omits the training-loop dispatch prerequisite. It describes ContextPilot-specific rollouts and actor-update wiring without explaining that dataset agent_name or a configuration override must select statelm_tool_agent; the inherited default is the generic single_turn_agent. The transformed training dataset is outside the frozen evidence. General no-execution and missing-data qualifications do not disclose this architectural alternative.

Unblock through a separately authorized correction route: qualify the runtime training route and public synthesis with the selector, fallback and unresolved dataset selection; align the boundary narrative and dependent assessments, then independently reverify. Alternatively establish selection from frozen evidence. This does not require executing training or acquiring data merely to state the condition. This report authorizes no repair or recovery. Historical holds remain evidence; this new hold is independently established against the exact fresh bytes.

Identity and method

  • run-id: AAS-2026-10-19-contextpilot-0826e2df5b2f-01.
  • method-commit: 3e2491c83a4613d410511af8c3cfde30f4b67689.
  • source-revision: 782cbb6611fb610c4cf6fafda6022b7e89cae191.
  • Source identity: https://github.com/Tencent/ContextPilot.
  • W: /media/zby/2b5e98b6-baf0-4e97-90d2-03321ea23372/outputs/commonplace-worktrees/batch-23-contextpilot.
  • S: W/related-systems/Tencent--ContextPilot.
  • P: W/kb/agentic-system-analyses/retained/tencent--contextpilot.
  • R: W/kb/agentic-system-analyses/state/AAS-2026-10-19-contextpilot-0826e2df5b2f-01.
  • L: /media/zby/2b5e98b6-baf0-4e97-90d2-03321ea23372/outputs/batch-23/contextpilot.output.
  • Preparation: sibling batch-23-contextpilot.preparation.json; ready, command prefix W/.venv/bin.
  • Auditor: independent Pi session 01a15553-6993-7087-8139-123237c86bac; observed PI_PROVIDER=openai, PI_MODEL=gpt-6.1-sol, PI_REASONING_LEVEL=medium. No delegation.

Followed the commission and audit procedure. Read the workshop README, workshop/text contracts, analysis collection and applicable set/runtime/synthesis contracts. Read the overview, synthesis and three verification members completely; selected boundary/runtime passages and direct frozen-source dispatch/configuration bodies for the consequential discrepancy. This is a bounded audit, not a second analysis or full memory-profile reassessment.

Current identity checks

Independently recomputed all intake-listed hashes: 1,020 source files, 13 publication files and 170 run files matched, as did preparation, publication effect and external log. P contains exactly its 13 expected members. The completed R/effects/publish.json names this run, source and destination; its new map equals the intake publication map, with expected incumbent absent. 000037-publish and the installed verification version were inspected. W HEAD equals the method commit; S HEAD equals the source revision and source Git status is empty. W status shows only its untracked publication directory. Preparation commit/path-prefix and overview identities agree.

Exact binding SHA-256
Preparation a15274b3df8ddf93dd17654bde2b1c03954ab2e76c24d1b8245ad7fdec8db788
Publication effect 9dc671cc82fd96840b4ef8e2fe5f059c96143ff7ad8ed23dc8d6dcb0cd79c4aa
External log L 6eb8f5c7fee38dbd3f20666bc5f6e672510f24158b4561aa1e16840f0486bb80
P/ARTIFACT.yaml a3d9788df8357a625bcb105c2045219fb9a39c8a7232a7957a9082cbf2e50c8c
P/runtime.md eaef3cdd36b6afc2bf401963a65bea5aa146a6356a85c37994207b041b5fca69
P/synthesis.md 78823bc3c69910b2fb3260ea2a5164c430b80dc3bfa8870bce0dc3e8476dfb8e

The complete maps and exact paths remain in intake cases.contextpilot. These checks bind publication, not integration readiness or every worker obligation.

Consequential source finding

The historical accounts supplied the question, not its answer. Direct inspection established this chain at S:

  1. train/verl/experimental/agent_loop/agent_loop.py:1006–1008 fills missing batch agent_name with config.agent.default_agent_loop. At 1080–1121, the worker selects that registry entry, instantiates it and calls its run.
  2. train/verl/experimental/agent_loop/statelm_agent_loop.py:451–452 registers the ContextPilot loop as statelm_tool_agent.
  3. train/verl/trainer/config/rollout/rollout.yaml:207–208 sets the default to single_turn_agent. ppo_trainer.yaml:8–36 imports that rollout configuration. The complete launcher-selected train/examples/sglang_multiturn/config/statelm_tool_agent_grpo.yaml inherits ppo_trainer without overriding the selector.
  4. The complete train/sh/run_qwen3-8b_longbenchv2.sh supplies tools, ContextPilot/partial-rollout flags and the estimator, but no default-loop override. It consumes external transformed Parquet files. P/boundary.md explicitly retains the separately downloaded training-data gap.
  5. train/verl/experimental/agent_loop/single_turn_agent_loop.py:26–64 registers the fallback and performs single-turn generation, not the ContextPilot tool/context-management loop.

This establishes a condition, not that actual training selected the wrong loop, that the dataset lacks its selector, or that the ContextPilot learning implementation is absent. ContextPilot and partial-rollout flags do not themselves establish the class selection.

Affected exact publication:

  • P/runtime.md:43–45 connects the training launcher to context-management snapshots and updates without the prerequisite. RT-RTE-training-rollout at 187–202 names the StateLM owner and lists ContextPilot/partial-rollout configuration as selection, but omits loop dispatch.
  • P/synthesis.md, “From rollout experience to a changed actor,” carries the branch/snapshot/update account without the selector or generic alternative. Its limitations withhold execution, exact models, data reproduction and downstream deployment, not this entry condition.
  • P/boundary.md:35–47 includes training snapshot consumers and says the recipe was traced through reward/update recovery without identifying conditional dispatch.
  • P/report-verification.md accepts the runtime/training chain and declares no blockers; its omission search addresses synthetic cleanup and alternate inference callers, not this training selector. Synthesis verification accepts the dependent public account. Neither acceptance supplies the missing evidence.

A search across every published Markdown member found no agent_name, default_agent_loop, single_turn_agent or statelm_tool_agent. The missing qualification therefore cannot be recovered elsewhere in this set. The synthesis contract requires preserving conditions that change whether a mechanism operates; an audit-only qualification cannot repair its absence.

Historical coverage and retained public limits

Read both intake-listed audits and the archived review linked by systems.md. All address the same source revision; a newer cutoff does not explain loss of its explicit dispatch condition. Their older traces, reset lineage and conclusions were not freshly certified here. The archived detailed result and depth investigations were not reopened.

The new public account preserves useful distinctions: per-sample stores rather than established cross-task autobiographical reuse; full-tool fallback and alternate base inference; mechanical review versus relevant semantic checking; original-history retention versus restoration availability; sampled-word compression and best-effort token windows; optional merge rather than default execution; descendant reward versus individual-action merit; conditional trainable synthetic cleanup; and wired updates versus exercised behavior or benefit. It additionally carries its verifier's selective recovery inventory, including automatic highest-ranked-chunk delivery, in public limitations. These are favorable coverage observations, not independent source certification of every statement.

Historical numerical-gradient/checkpoint details, reflection interpretations and all memory axes were not exhaustively compared. No disposition on those unsettled differences is needed to decide this hold. Further reconstruction stops at the established dispatch defect.

Bounded operational evidence

Parsed settled outer message_end subagent results in L, not repeated stream snapshots. There are 13 analytical worker results plus the diagnostic; all report exit zero. Sampled boundary, runtime and report-verification nested messages. Calls below were paired with results by exact ID inside the same worker; message indices are zero-based:

L line / worker Nested messages and exact call ID Established action
765 / 000003-boundary 14→15, call_pRR3JZVrTMsIuPg4KFfCYevA\|fc_0515a0a65cb57ba6016ad619fad61487d2a587ff20d540472c Read S/README.md; result isError: false.
1151 / 000005-runtime 21→22, call_4dn8R5m7QhIdNiL4DYbyazIR\|fc_0687e270c4ecda2f016ad61adfc52887d2ad9b6e60726be3b2 Read S/infer/src/contextpilot_fsm.py, offset 1, limit 175; result isError: false.
2844 / 000022-report-verification 50→51, call_OJqajN9ZVTw5hniLBgmkX4Av\|fc_0f8f1bd538f955a7016ad61f02d60887d2a59697afaddc2eca Read S/infer/src/contextpilot.py, offset 580, limit 210; result isError: false.
2844 / same verifier 95→96, call_fyStK0VkoVHx2UvLWkIgKvIX\|fc_0f8f1bd538f955a7016ad61f62c11087d2b7136cd2e86ca6f2 Successful write to its handout outputs/verification.md; later 99→100 successfully edits that same output.

The last row is evidence of actual output mutation, not an unexecuted suggestion. The submitted verification hash is the installed report-verification version. Source-read samples establish some independent inspection, not inspection of the missing dispatch condition. No new operational violation was established by these bounded samples. Complete invocation recovery, all executed shell effects, every write's authorization, indirect legacy exposure, unsampled worker compliance and current writer idleness were not certified. Publication success and zero worker exits do not discharge those obligations or clear historical incidents.

Validation and limits

From W with PATH=W/.venv/bin:$PATH, ran only read-only validation, not a run/engine command:

commonplace-validate kb/agentic-system-analyses/retained/tencent--contextpilot
VALIDATION SUCCESS
Target: kb/agentic-system-analyses/retained/tencent--contextpilot
Scope: kb/agentic-system-analyses/retained/tencent--contextpilot
Files analysed: 1
Text files: 0
Diagnostics: 0 failures across 0 subjects; 0 warnings across 0 subjects

The count denotes one directory artifact, not twelve separately counted documents. From the same prepared environment, report validation returned:

commonplace-validate /home/zby/llm/commonplace/kb/work/agentic-system-reruns/audits/batch-23-contextpilot.md
VALIDATION SUCCESS
Target: /home/zby/llm/commonplace/kb/work/agentic-system-reruns/audits/batch-23-contextpilot.md
Scope: /home/zby/llm/commonplace/kb/work/agentic-system-reruns/audits/batch-23-contextpilot.md
Files analysed: 1
Text files: 1
Diagnostics: 0 failures across 0 subjects; 0 warnings across 0 subjects

Deterministic success does not resolve the semantic hold.

Only this assigned report was authored. No analysis/source/state changes, delegation, commits, run effects, integration, acquisition, target execution, cleanup, termination or push occurred. No analyst feedback was sent. Other reports and unpublished timeout cases were outside scope. Actual dataset selection and target benefit remain unknown; no missing source prevents stating the established condition.