Capability-gate host receipts round 2 at main eb5ca78a (supersede the first recording; GPT-6 review agree) + dashboard gate row - #442

Scope

• What this PR changes:
  • It adds three round-2 host receipts for the PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381 Codex capability gates on nativestack-5975wx-20260925. Each supersedes its first recording, which carries the GPT-6 review's needs_changes and was published in Capability gate: retain scored results, enforce tool identity and directory-change forms (GPT-6 review repair); first host receipts with that review #440.
  • It adds a grand-dashboard gate row for them.
• How they were recorded:
  • Each receipt was recorded with scripts/host_receipts.py record --supersedes <first receipt id>, from a detached worktree at published main eb5ca78a (Capability gate: retain scored results, enforce tool identity and directory-change forms (GPT-6 review repair); first host receipts with that review #440).
  • Recording re-ran tools/capability-gate/run_gate.py live.
• Base commit: eb5ca78a (Capability gate: retain scored results, enforce tool identity and directory-change forms (GPT-6 review repair); first host receipts with that review #440) for the recordings; the branch merged main b26add90 (GPT-6 gateway lane on the all-engines OmniRoute instance: static headers and a one-slash model #439) before its hot-file commit
• Lane: lane:foundation
• Paths touched:
  • the three -2 receipt files under evidence/hosts/nativestack-5975wx-20260925/;
  • observability/grand-dashboard/state.json: the new gate row codex-capability-gate and recorded_at_utc;
  • manifests/evidence.json: registration only;
  • the regenerated reports.

SOTA sources

• promptfoo 0.123.1 (tag commit 34f74d34): the openai:codex-sdk provider, evaluator, assertions and hooks, and the results sanitizer. The harness README and its decision record (Capability gate: promptfoo codex-sdk harness for the Codex jcodemunch, ai-memory and M13 gates, reconciled with native OTel #433, amended in Capability gate: retain scored results, enforce tool identity and directory-change forms (GPT-6 review repair); first host receipts with that review #440) cite them line by line.
• openai/codex rust-v0.157.1 (tag commit ac0e23e5): the per-server enabled and approval keys (decision F7), and the approval-refusal error at codex-rs/core/src/mcp_tool_call.rs#L1612.
• Loki v3.7.8 (tag commit 09e6ce2f): query_range bounds and label filters on structured metadata.
• The components exercised:
  • jcodemunch-mcp 1.108.319;
  • ai-memory 2.4.1, the host daemon;
  • context-mode 1.0.169;
  • serena_agent 2.0.0.dev0 from oraios/serena at c6fbd1c5, the landscape code-navigation winner pin. This is recorded as a limitation in the M13 receipt, whose component is context-mode.

Evidence-class table

Each receipt claim quotes only that recording's own verdict line. The recording script kept a receipt only if its result was pass, its excerpt started with the quoted line, and the line carried a 64-hex results field. All three were kept.

Each receipt's limitations state what PASS means in the harness at eb5ca78a:

• the argument fields actually checked;
• the directory-change forms, a syntactic list;
• sentinel retention;
• shell visibility and Loki reconciliation;
• for M13: top-level-only scope, overlap, wrong-root indexing and the Serena build;
• for ai-memory: the pin mismatch and the empty-hit boundary;
• for jcodemunch: route left unqualified.

Local commands run

$ python3 scripts/host_receipts.py record ... --supersedes <first receipt id>   (x3, from detached worktrees at eb5ca78a)
KEEP m13 pass | KEEP ai-memory pass | KEEP jcodemunch pass   (recording 23:28:10Z-23:43:34Z)
$ node <re-score> tools/capability-gate/assertions.js <retained results.json>   (x3, counts only)
m13 260/260 identical; ai-memory 10/10; jcodemunch 10/10; m13 rows with own sentinel 52/52
$ python3 -m unittest tests.test_grand_dashboard
OK
$ python3 scripts/component_matrix.py --write
{"flip_rule_violations": 0, "rows": 32, "status": "written"}
$ python3 scripts/new_host_grand_list.py --write
{"status": "written", "layers": 32, "winners": 66}
$ python3 scripts/validate.py
{"components": 69, "hashed_files": 7373, "profiles": 4, "receipts": 159, "status": "passed"}
$ GIT_NO_LAZY_FETCH=1 python3 scripts/host_receipts.py validate
{"receipts": 182, "status": "passed"}
$ git diff --check origin/main...HEAD   # clean; pre-push registry gate: Ran 3 tests OK

Independent review

• Who reviewed. The reviewer is the one whose needs_changes the first receipts carry: gpt-6-astra at effort max, read-only codex exec, its own thread resumed.
• What it checked. It reviewed each -2 receipt at eb5ca78a:
  • the claim quotes only the recording's own verdict line;
  • the result, supersession and date are consistent;
  • the limitations match the harness;
  • it re-applied assertions.js to the retained results itself, as an independent replay.
• Replay results. 260/260, 10/10 and 10/10 component verdicts, with the gate, disabled and prompt/wrong-root tallies all matching. Retained hashes matched 3/3, and M13 sentinels are bound in 52/52 rows with 52 unique values.
• Findings. No P0–P3 findings.
• Recorded verdicts. It then recorded codex_lane agree on each receipt itself, with scripts/host_receipts.py review under its own identity, which differs from the recorder's.
• Coordinator-supplied inputs. The Loki, model and installed-Serena observations were supplied by the coordinator. The reviewer counted MCP calls independently from the retained results and matched the supplied totals.
• Round 1. The first receipts keep their needs_changes reviews (Capability gate: retain scored results, enforce tool identity and directory-change forms (GPT-6 review repair); first host receipts with that review #440). Superseded at the same version, they support no status.

Host evidence

• Where recorded: host executions on nativestack-5975wx-20260925 (linux-wsl2-x86_64) at eb5ca78a, from 2026-09-27T23:28:10Z to 23:43:34Z. Historical receipts; they do not certify another host.
• Run conditions:
  • Main checkout. The jcodemunch sessions run in the main checkout, as the harness requires. It stayed at c8362c02, and the two fixture files (scripts/host_receipts.py, tools/sota-convergence/landscape-sweep/build_args.py) do not differ between c8362c02 and eb5ca78a.
  • Codex home config. It was last modified at 2026-09-27T21:42:49Z, before the window. The round-1 SocratiCode switch did not recur, and SocratiCode shows 0 tool results.
  • Retained results. They stay host-private: files 0600, directories 0700, never overwritten. They are not published.

🤖 Generated with Claude Code