Correct the native mapping checkpoint after implementation review round1. You remain sole implementer antigravity/gemini-3.8-flash-high, full-access/default. No delegation/review/commit/push/live Herdr/config changes.
Same repository main HEAD0c9991439309b1ce819abc4eb007664b872cc45b. Both operative freezes and approval remain unchanged: base docs/plans/2026-10-17-two-surface-orchestration.md @58d48cea483804b22767b381c6edd8274fe6ad44; amendment docs/plans/2026-10-17-herdr-native-pilot-critic-amendment.md @b99e212a55b20757eeda8605db4edcc59aad6107. Verify them. Read original worker brief implementation/native-mapping-worker-brief.txt and original correction brief implementation/native-mapping-correction-brief.txt. Preserve coordinator's later example corrections in current files.
Allowed edits exactly SAME SIX files beneath docs/validation/two-surface-orchestration/2026-10-17/: setup.md,pilot.md,acceptance.md,implementation/preflight-receipt.json,implementation/human-pilot-brief.md,implementation/native-mapping-application-report.md. External scratch /tmp/skill-lens-herdr-native-mapping-implementation-20261017/. No other file changes. .path untouched, historical reports and evidence untouched. P01-P09 stay PENDING; Phase C unstarted.
Round1 MiniMax reported changes_required with two reproduced blockers:
B1 P03 WORKER_TOKEN only defined/prose; fixed literal worker agent name, no complete literal executable implementer prompt/send/result binding. Need unique regex-valid worker name (worker-hex12), exact input path+initial SHA+token in literal complete independent WORKER_PROMPT (with full constraints, output authority, no delegation/review/commit/push, full implementer schema+identity channels), actual prompt send, own submission/result capture and complete token+path+hash response acceptance, independent edit comparison. Native fields N/A with controls, served identity unverified. No automatic prior worker session resume.
B2 P03 post-implementation reviewer wave currently only TOKEN3/4 and prose. Supply FULL executable preparation+dispatch+collection for both fresh Pi critics: unique regex-valid names, nonempty distinct ready new pane IDs, exclusive empty dirs, exact approved model/native flags, actual separate process/session identity/readiness, two complete literal independent prompts with verbatim critic guard and REVISED path+SHA+token, both submission jobs launched before parent collection, each exit/log result captured, complete report recovery/binding and clean content fingerprints. A concrete reusable snippet fully defined inside the copy/paste brief is acceptable only if invocation/rebinding is explicit and nothing is left to invent; do not add a repo script/framework. Do NOT merely say repeat P02. Keep all existing pre-edit P02 safety.
Round1 DeepSeek initial and one completion recovery ended narration/malformed tool output, no verdict; both terminal without children/questions, zero files changed. NOT PASS and not extra substantive review rounds. Do not manufacture DeepSeek agreement. Fresh required both-reviewer tasks will follow corrections, with inline context to avoid retrieval failures.
In same correction resolve useful MiniMax suggestions:
- P02 each literal EFFECTIVE_MODEL tells native runtime/interaction N/A and records actual controls and requested/configured/confirmed-or-unverified.
- Nonempty assertions for every parsed pane ID, each actual available foreground shell verified before start, distinctness against coordinator and other wave panes; STOP on failed split/start/prompt rather than empty ID/focus fallback. Handle failure logs, preserve uncertain delivered state; no blind resend.
- acceptance application section cite own native-mapping-application-gates logs:1048pass1skip91.11%,mypy37,Ruff clean,format216. Distinguish baseline215/previous205/197/187/184 counts as separate historical runs, never reconcile them into one. Final post-correction gates not yet run; coordinator will run.
- Inline pointer from older acceptance setup/amendment historical pending wording to adopted section.
- Reinstate pilot P02 no mixing T3 critics/wrappers/kind substitution within pilot.
- Report note coordinator gates already run versus future post-correction checks; no claim tested P09 or live readiness. Update mappingApplication task provenance current correction but preserve historical.
Additional coordinator concern: worker and fresh critic reports must use unique per-dispatch token + exact path+SHA; original-repo external helper content fingerprint must remain (not replaced by Git status-only). Save/compare every wave and stop on differences; helper exit code alone insufficient. P04 actual assistant complete ordered payload; Pi NO write fallback; stale token/unresolved liveness reject, human-only startup/restart continuation and protected cleanup stay intact. Do not weaken current text.
Do needed static checks (JSON/links/verbatim guards/name syntax/shell snippets) using external scratch, no live execution. Return seven council implementer fields only; no open-ended narration. Final source checks and independent reviews owned by Sol.
Full MiniMax report for context:
EFFECTIVE_MODEL: requested/configured = T3 Code `delegate_task` reviewer, provider instance `minimax` (mcode), model `MiniMax-M3.1-Flash-Preview` with thinking effort `xhigh`, `runtimeMode=full-access`, `interactionMode=plan`, role=review; upstream served-model/family identity UNVERIFIED (catalog + T3 receipts prove the configured selection only, no provider-confirmed identity). I did not inherit or reuse any other model's identity. Session: `mvs_35a2805c87fb48bda3b78b81282a894b`.

VERDICT: changes_required

BLOCKERS:
1. `docs/validation/two-surface-orchestration/2026-10-17/implementation/human-pilot-brief.md:212` + `:221` — P03 generates `WORKER_TOKEN` but never binds it to anything executable: the launch uses the fixed literal name `worker` (`herdr agent start worker --kind agy --pane "$PANE3_ID" ...`), no literal implementer prompt text exists (line 227 is descriptive: "Supply WORKER_TOKEN ... in the implementer prompt"), and the token appears in no name, path, or envelope field. Expected: base plan §4.2 "Every worker receives a unique per-dispatch token ... unique agent name" plus "A report envelope echoes the token and target ... Accept a report only after matching both" — the same brief already satisfies this for critics (`critic-qwen-${TOKEN1}` at :106, literal `PROMPT_QWEN` at :131-154). Evidence: `grep -n "WORKER_TOKEN|agent start worker" human-pilot-brief.md` → token defined only at :212 and referenced only in prose at :227; name literal `worker` at :221. Impact: the operator has no defined string to send and no acceptance key for the implementer report; `worker` also breaks the per-dispatch unique-name rule the same file enforces for critics.

2. `docs/validation/two-surface-orchestration/2026-10-17/implementation/human-pilot-brief.md:244-247` — the post-implementation review wave is not executable as written. `TOKEN3`/`TOKEN4` are generated, but the section supplies no agent names, no regex check, no session-directory creation (`mkdtemp`/`mkdir` without `-p`), no pane split/parse, no `agent start`/`agent get` readiness, no literal prompt with the revised `TARGET_SHA256`, no background submission, and no exit-code capture — unlike P02 at :96-208. Expected: the handoff criterion "complete independent critic prompts" and base plan P03 ("handoff to critics"), and acceptance.md:141 claims "complete independent supplied briefs defined before submission" for this brief. Evidence: brief:244-247 is the entire wave definition; contrast P02 :96-208. Impact: P03 cannot be executed or evidenced without the coordinator improvising, which is exactly the class of gap the earlier coordinator defect round was meant to close.

SUGGESTIONS:
- `implementation/human-pilot-brief.md:149` and `:174` — the critic `EFFECTIVE_MODEL` line requests `runtimeMode, interactionMode` but the prompt never tells the critic to mark them `N/A` and record native controls instead (amendment §3 table row "T3 permission/interaction fields | N/A on Herdr"). Only P04 (`:258`) carries that instruction, so the P02 pre-edit wave can yield a meaningless or omitted field.
- `implementation/human-pilot-brief.md:97-100` — add an explicit non-empty assertion on each parsed pane ID. `test "$PANE1_ID" != "$PANE2_ID"` catches two empty parses but passes when exactly one `pane split` yields empty JSON, deferring failure to `agent start`.
- `acceptance.md:120-149` — the new checkpoint section does not cite its own four-gate logs (`implementation/native-mapping-application-gates/`). The current run (1048 passed / 1 skip / 91.11%, mypy 37, ruff check clean, **216** formatted) is verified in those logs but recorded nowhere in the six files, while acceptance.md:76-80 carefully distinguishes 197/187/184. Five distinct ruff-format counts now exist (184, 187, 197, 215, 216) and only four are labelled as separate runs.
- `acceptance.md:101-105` — "Setup approval update" still asserts "exact replacement selection, controls, operative approval/freeze, human startup and P01-P09 remain outstanding" with no inline pointer to the superseding final section, unlike the explicit supersession note at `:103-105`. Add "superseded by the final section" so a reader cannot mistake it for a live blocker.
- `pilot.md:31` — the P02 row dropped the prior clause "never mix T3 critics into pilot, install wrappers, or substitute kinds without amendment". That constraint still lives in `human-pilot-brief.md:73` and frozen §4.4, but the matrix itself no longer carries it.
- `setup.md:35` and `:51` — `which cmux` / `which mastracode` use the deprecated `which`; `command -v` is the modern form. Cosmetic only.
- `implementation/native-mapping-application-report.md:107` says "Coordinator will execute the final four gates after inspection"; the coordinator's run (13:18, 216 files) postdates this report and is not noted as such inside it. Administrative only.

FILES_AT_RISK:
- `docs/validation/two-surface-orchestration/2026-10-17/implementation/human-pilot-brief.md` — primary; both blockers are in the executable handoff (P03 dispatch at :210-247).
- `docs/validation/two-surface-orchestration/2026-10-17/acceptance.md` — record completeness of the checkpoint's own gate evidence and one stale-section pointer.
- `docs/validation/two-surface-orchestration/2026-10-17/pilot.md` — P02 row lost a standing no-substitution clause.
- `setup.md`, `preflight-receipt.json`, `native-mapping-application-report.md` — no in-scope defect found.

CONCEDED:
- Both frozen contracts preserved. Base `58d48cea483804b22767b381c6edd8274fe6ad44` + amendment `b99e212a55b20757eeda8605db4edcc59aad6107` are cited in all six files; the amendment's operative sections 1–7 are byte-identical to `native-critic-amendment-reviewed-draft.txt` (independently verified: extracted spans equal, 11914 chars each). Frozen plan and amendment untouched.
- Exact approved routes/permissions reproduced verbatim: `provider=anymodel` / `model=qwen/qwen3.8-max` `--thinking high`; `provider=workbuddy` / `model=deepseek-v4.1-flash` `--thinking high`; `--tools read,grep,find,ls --no-mcp --no-extensions -e .../herdr-agent-state.ts --no-skills --no-context-files`. Cross-checked against installed static help: `pi --help` documents `--tools read,grep,find,ls` as a supported example (line 120), "Keeps MCP tools unless an entry starts with mcp__" (line 38, justifying the `--no-mcp` rationale in setup.md:100), and `--no-extensions ... (explicit -e paths still work)` (line 44). Upstream identity/family stays unverified everywhere; no Grok/XKiro or unapproved fallback appears in executable instructions.
- Empty fresh per-dispatch conversations/processes: `tempfile.mkdtemp` inside the owned artifact root plus `mkdir` without `-p`, no `--continue`/`--resume`, distinct dirs per critic. The resolved "directory reuse" defect is genuinely fixed.
- Distinct ready panes: the JSON envelope `["result"]["pane"]["pane_id"]` used at brief:97-98 is confirmed by installed `herdr --skill` (lines 123 and 181: "Read the new pane ID from `.result.pane.pane_id`"); the name regex `[a-z][a-z0-9_-]{0,31}` is confirmed at skill line 57, and `critic-qwen-<hex12>` (24) / `critic-deepseek-<hex12>` (28) both fit. `--current/--direction/--cwd/--no-focus`, `agent prompt --wait --timeout`, `agent read --source recent-unwrapped --lines`, `agent get`, `agent send-keys`, `status --json`, `session stop|attach`, `pane split` all verified against installed static help.
- All submissions before collection: both `herdr agent prompt ... &` are backgrounded before either `wait`, with separate exit codes and separate `.result` files (brief:185-196), plus explicit no-duplicate-send reconciliation (:198-199) consistent with `agent prompt --help`'s own warning.
- Token+path+SHA binding present in both critic prompts and required for acceptance (brief:146-148, :205-208).
- Pi complete-report capture or STOP with **no** write fallback is preserved in all three documents (brief:260-265, setup.md:104, pilot.md:19), and the amendment's no-tool-widening rule is explicit ("Do NOT grant write or shell tools to rescue capture"). P04 correctly requires the generated response, not the echoed prompt.
- Native restrictions are documented as a tool/context restriction, not a filesystem jail (setup.md:105-106, pilot.md P05, brief:275). Bounded owned-stop/liveness proof retained (brief:283-292). Human-only lifecycle retained, `herdr server stop` explicitly prohibited (setup.md:125, brief:300-304). Repository read-only + external-artifacts-only output retained (brief:67-74, setup.md:180-183).
- Repo stays read-only during the pilot and this checkpoint: `git status` shows only the five allowed tracked docs/JSON modified plus the new untracked report; `.path` untouched; no Python/test/config/registry/process-doc change; `git diff --check` clean.
- P01–P09 remain PENDING with no paper pass: 11 `PENDING` occurrences in pilot.md, 9 `"PENDING"` in preflight-receipt.json, all nine rows `Unexecuted` in acceptance.md. Phase C unstarted (`surfaces/t3.md` / `surfaces/herdr.md` "Missing"). No trial-ready or supported claim anywhere.
- The external content-hash helper genuinely exists and matches the brief's contract: `/tmp/skill-lens-herdr-execution-20261017-cg7iv351/snapshot_tree.py` implements `save`/`compare`, hashes every tracked and non-ignored untracked file plus symlink targets, prints `changed_paths`/`head_unchanged`/`status_unchanged`, and does not signal differences via exit status — exactly as brief:218 and :242 state. Root comes from `cwd`, which the brief's `(cd "<repo>" && ...)` supplies.
- Handoff is paste-ready: the fenced coordinator block was extracted and its 58 command lines parse cleanly under `bash -n` (the whole block is prose-for-the-agent, not a script, so a block-level parse error is expected and not a defect). Both critic prompt assignments are quote-safe — 1288 chars each, zero stray `"`, backtick or backslash, and the only expansions are `${TOKEN1}` / `${TOKEN2}`.
- Fixture is real and matches every recorded hash: `/tmp/skill-lens-herdr-execution-20261017-cg7iv351/synthetic-checkout/fixture.txt`, SHA256 `462b045a…7014e`, content `pilot status: pending`. Pilot-artifacts directory is currently empty, consistent with an unrun pilot.
- Source gates verified from their own logs (source health only, never runtime): 1048 passed / 1 expected case-insensitive-FS skip / 91.11% coverage, mypy clean on 37 files, Ruff check clean, **Ruff format 216 files already formatted**, with synthetic HOME/XDG and external cache/coverage — matching the reported figures exactly.
- All local link targets referenced from the six files resolve on disk; the three JSON evidence files parse; `native-critic-amendment-execution-receipt.json`, both round-2 reviewer reports and the receipts/window files are present and correctly labelled as preflight/amendment reviews, not final migration acceptance.
