REQUEST_TOKEN: omp-external-doc-final-20261022T004510Z

     EFFECTIVE_MODEL:

     - Requested: opencode/minimax-m3.1-flash-free, default settings,
       Plan mode (per the review request and the migration's one-
       reviewer exception in docs/process/omp/README.md:33).
     - Configured: harness-declared "MiniMax M3.1 Flash Free", variant
       max (as displayed in the submission UI); Plan/read-only critic
       mode, no tools that write.
     - Evidence: harness/system declaration only. Upstream served-
       model identity is unverified — no provider handshake, no
       independent confirmation. Consistent with docs/plans/2026-10-
       22-omp-herdr-external-control.md:8,99,128.
     - Transport: this review runs through the migration's external
       OpenCode document-review relay (request token prefix omp-
       external-doc-final), not through a Herdr worker. See SUGGESTION
       1 — the frozen contract distinguishes "the separately
       authorized review transport" from the "bootstrap relay" and
       asks that any fallback not be silent.
     - Mode: Plan mode; no repository writes, no delegation, no
       background tasks, no Herdr commands, no installed skill/
       settings changes.

     VERDICT: pass

     Review basis: frozen contract docs/plans/2026-10-22-omp-herdr-
     external-control.md @ 988d3917890ddf56b48b950afa2c162dd84c8c4e (
     read in full, §"Implementation contract" §1–§4, §"Verification
     and acceptance"), the coordinator-supplied diff base
     321e5693f9d8e528bb350e465fab80655d3d1922 → working tree, the
     evidence bundle (implementation.diff, baseline-text.json,
     document-validation.json, gates.txt, pre-final-review.json), and
     the current on-disk contents of all five changed files plus the
     historical evidence record. Gates were read, not run.

     BLOCKERS: none.

     What I verified as conforming (each traced to a contract clause):

      1. Native Tern coordinator, retained identity check — docs/
         process/omp/herdr.md:9 keeps the README "Start directly"
         openai-codex/gpt-6.1-sol check; no silent fallback (contract
         §1 opening). The route is Sol-native-in-Tern with Herdr
         hosting only workers (herdr.md:3,40; README.md:17; planning-
         loop.md:54; implementation-loop.md:40).
      2. Named existing session/socket verification; missing/ambiguous
         stops — herdr.md:13-32: herdr session list --json
         enumeration, select an authorized existing named session (
         recording default explicitly), require running + socket_path,
         test -S Unix-socket check, herdr --session <verified> status
         server --json cross-match, session:null/default normalization
         only when the discovery row has name:"default", default:true
         and the identical socket, and an explicit stop on missing/
         stale/incompatible/ambiguous identity. Root --session can
         create, so discovery must prove an existing running entry
         first; no create/attach/repair/fabricated context/historical
         IDs. Same --session prefix pinned on every later command.
      3. Narrow, user-approved skill exception with no skill/settings
         edits — herdr.md:11 names the conflict (inside-Herdr/
         HERDR_ENV gate, caller-relative defaults), names the
         amendment + adapter as authority, keeps all non-conflicting
         skill safety rules, and disclaims system/runtime overrides,
         approval widening, installed-skill/settings edits and other
         projects; README.md:17 mirrors it. Matches contract §1.1
         verbatim in substance.
      4. Explicit pane IDs throughout — herdr.md:34 (no --current,
         inherited HERDR_PANE_ID, focus, sidebar order or agent-name
         substitutes), base-pane inspection templates with explicit
         IDs (:42-45), split with --pane <base> … --no-focus (:50),
         returned .result.pane.pane_id recorded with ownership before
         startup (:53), explicit IDs in prompt/get/read/wait (:56-59,
         65-66), and cleanup "explicit pane IDs and the same verified
         session prefix" (:81).
      5. Timeout on every wait — startup --timeout 30000, prompt --
         wait --timeout 120000, agent wait --timeout, pane wait-output
         --timeout, plus the blanket rule "Every wait must have an
         explicit positive finite --timeout in milliseconds … do not
         invent undocumented maxima, wait indefinitely or increase
         timeouts to disguise the known defects" (herdr.md:62).
      6. Inspect-before-resend — herdr.md:77 covers timeout,
         agent_prompt_stalled, blocked/unknown, connection loss; no
         blind resend or duplicate worker; timeout ≠ cancellation/non-
         delivery; agent_not_idle = incomplete readback, not resend
         authority. Report-file fallback rules retained (:79).
      7. Guard, ledger, roles, review and acceptance rules unchanged —
         the diff touches only the routing sentence in each of the
         three route docs and only the Herdr section + status sentence
         in the README; README's guard block, review policy, roles,
         ledger and completion-evidence sections are byte-for-byte
         untouched (independently confirmed against the diff hunks and
         the current file; the guard text also matches the guard in
         this very request). No new permanent tests, no product/source
         changes.
      8. Consistent route across four docs — README §"Herdr workers
         from native Tern", planning-loop.md:54, implementation-loop.
         md:40 all state: Sol native in Tern, external connection
         gate, missing/ambiguous connection stops that route (not
         merely being outside Herdr), standalone routes retained when
         Herdr is not requested, never a silent switch to native
         tasks/relays. No relaunch-inside-Herdr instruction remains in
         executable guidance in the four documents (verified by
         reading each file and grepping the tree; no stale #when-
         running-inside-herdr / #establish-the-real-caller / #live-
         check-still-required anchor is referenced anywhere).
      9. Dated append preserving history and full reference — docs/
         plans/2026-10-20-omp-herdr-dispatch.md:38-46: original text,
         authorization, review history and Reference:
         2656a6a1efb687b0d23c9679fdc2222cc18c07b0 preserved verbatim;
         the append cites the amendment's full hash
         988d3917890ddf56b48b950afa2c162dd84c8c4e, limits supersession
         to the inside-Herdr caller/launch/targeting setup, and states
         the amended procedure (native Tern start, named session/
         socket verification, explicitly authorized base pane, one
         owned worker pane, explicit IDs, bounded waits) while
         retaining the original acceptance bar.
     10. Failures stay unresolved; no live acceptance claimed —
         adapter and README status sentences both read "external
         dispatch documented/unverified; end-to-end acceptance is
         unsatisfied" and cite the Mini busy-detection failure and the
         ~6.91 s early full-TUI waiter return. I checked both claims
         against the historical evidence record docs/validation/omp-
         coordinator/2026-10-20/herdr-dispatch.md (:320 for the 6.91 s
         early return; :225-233 for the esc stop/screen-rule mismatch)
         — both are accurate. herdr.md:83-89 and the append state that
         documentation implementation does not authorize a smoke, that
         a report recovered after a premature waiter return does not
         pass wait-correctness, and that a reviewer-only smoke cannot
         establish implementation-worker capability. The historical
         evidence file was not rewritten (still contains "Human step
         needed for live acceptance") and is superseded by reference,
         not edit.
     11. Scope discipline — exactly five paths changed (document-
         validation.json changed_paths; diff hunk headers); the pre-
         existing evidence edit, .path and the 2026-10-19 draft appear
         only as preserved (pre-final-review.json status  M …/herdr-
         dispatch.md, ?? .path, ?? docs/plans/2026-10-19-herdr-omp-
         coordinator.md) and are absent from the diff. Gates log: 1082
         passed / 1 skipped / 91.16% coverage, mypy clean (38 files),
         ruff check clean, ruff format clean (278 files) — read, not
         rerun.

     SUGGESTIONS (non-blocking; none requires a protected-policy
     change):

     1. Name the final-review transport in the implementation receipt.
        The frozen contract's post-implementation acceptance says "Use
        the separately authorized review transport; if Herdr is
        unavailable or cannot meet completion requirements, stop
        rather than silently use the bootstrap relay" (docs/plans/
        2026-10-22-omp-herdr-external-control.md:92), while herdr.md
        and the contract bar any live smoke. This review therefore ran
        on the external document-review relay, not a Herdr worker.
        Expected: one receipt line naming the transport, requested
        model, session id, exit status, that Herdr transport was not
        exercised, and why (no smoke authorized). This closes the
        "silently" condition without changing any rule.
     2. The implementation execution receipt is still owed. docs/
        plans/2026-10-22-omp-herdr-external-control.md currently ends
        with the planning/freeze receipt (line 133) and has no
        implementation receipt, while docs/process/omp/implementation-
        loop.md:82 requires one in the dated plan/execution record (
        expected vs actual launches, corrections and reasons,
        requested/configured identities and settings with upstream
        limits, report/log paths, elapsed time/tokens if exposed).
        Frozen-plan scope item 5 permits appending it "if needed"
        without changing the operative contract. Suggested location: a
        new dated section appended at the end of that plan; do not
        touch the operative contract or Reference:, and record the
        resulting new file digest.
     3. Cite the baseline in the validation receipt. document-
        validation.json asserts "protected_policy": "unchanged" and
        "original_smoke_plan": "preserved exactly" but does not record
        the baseline commit it compared against (the contract names
        6b77388003ca8c411b19452d94528344bb8e1007), the diff base (
        321e5693f9d8e528bb350e465fab80655d3d1922), or the manifest
        digest (
        c67fbb01e50b31d1a854891d60aed3cb72ca6a1791d6ff57c7a2116ee28e6e
        ae). Recording all three makes the claim independently
        auditable. (I verified the protected sections myself from the
        supplied diff and current file contents; this is a receipt-
        completeness point, not a defect.)
     4. One residual reading ambiguity in the 2026-10-20 plan (worth
        one clarifying clause). The append states "Only the caller/
        launch/targeting setup changes … The exact reviewer guard,
        ledger, independent-review and acceptance rules stay
        unchanged", while the preserved plan body still contains the
        pre-amendment route condition "Herdr-backed dispatch is
        selected … whenever OMP is run inside Herdr" (:29) and the
        inside-Herdr caller/setup steps (:21,22,25). The contract
        required that body to be preserved and limited the
        supersession to "caller/launch/targeting setup", so nothing is
        wrong as written — but a reader could take "acceptance rules
        stay unchanged" to keep the old route-selection condition
        alive, which conflicts with the amended rule (route keys on
        the user requesting Herdr, not on being inside Herdr).
        Suggested: one added clause in the append noting that the
        route-selection condition in the preserved acceptance bullet
        reads as "when Herdr dispatch is requested" for this OMP
        route. This is additive text in the new append, not a rewrite
        of the preserved history.
     5. Optional pane-discovery template. herdr.md:40-45 requires an
        explicitly authorized base pane ID and then inspects it, but
        gives no session-scoped command for discovering a pane ID (the
        old --current route is correctly banned). Consider adding a
        pane-listing template taken verbatim from installed herdr pane
        --help so an operator can obtain a base pane ID without
        relying on focus. Non-blocking: the current wording is fail-
        safe (explicitly authorized ID required, ask when ambiguous).
     6. Optional, later-authorized only. The README's "Tested external
        OpenCode review route" section is correctly untouched (it was
        outside the amendment's allowed edit scope), but it still
        reads as a generally available route, whereas contract §3 asks
        that bootstrap review evidence remain "explicitly historical/
        transitional". If a later authorized edit touches the README,
        one clause marking that route as the migration's document-
        review bootstrap and not the Herdr worker route would close
        the gap. Do not change it under this amendment.

     FILES_AT_RISK:

     - Changed by this implementation (reviewed, conformant): docs/
       process/omp/herdr.md, docs/process/omp/README.md, docs/process/
       omp/planning-loop.md, docs/process/omp/implementation-loop.md,
       docs/plans/2026-10-20-omp-herdr-dispatch.md (append-only).
     - Expected next touch, append-only, must not alter the operative
       contract: docs/plans/2026-10-22-omp-herdr-external-control.md (
       implementation execution receipt).
     - Preserved user work, must not be staged, reverted or rewritten:
       docs/validation/omp-coordinator/2026-10-20/herdr-dispatch.md (
       pre-existing modification), .path (untracked), docs/plans/2026-
       10-19-herdr-omp-coordinator.md (untracked).
     - Untouched by design: README guard/review/ledger sections, docs/
       process/omp/* non-routing rules, T3 workflow, installed herdr
       skill and settings, product source/tests.

     CONCEDED (limits of this review):

     - I reviewed the coordinator-supplied diff (
       321e5693f9d8e528bb350e465fab80655d3d1922 → working tree) plus
       the current on-disk contents of the five changed files. I did
       not run git, so I could not independently recompute the
       manifest SHA-256
       c67fbb01e50b31d1a854891d60aed3cb72ca6a1791d6ff57c7a2116ee28e6ea
       e, the per-file digests in pre-final-review.json, or confirm
       that the tree still matches them at review time; that evidence
       is coordinator-asserted.
     - No gates, tests, linters, formatters, link-checkers or runtime
       commands were executed; gate outcomes are read from gates.txt
       only. I did not re-derive the "33 links/anchors" count by
       machine; I verified the anchors used in the five documents by
       hand (all resolve to existing headings/files).
     - No Herdr commands, no session/socket discovery, no worker
       launch, no live smoke — none is authorized, and none was
       attempted. Historical Mini and full-TUI failures remain
       unresolved by this review.
     - I did not verify the protected-text comparison against baseline
       commit 6b77388003ca8c411b19452d94528344bb8e1007 by hashing; I
       verified it structurally from the diff and the current files (
       only routing sentences changed in the three route docs; only
       the Herdr section and status sentence in the README).
     - Upstream model identity remains unverified (harness declaration
       only). This real document review is not Herdr lifecycle
       certification, not end-to-end acceptance, and not a smoke.
     - No files were created, edited, moved, deleted or renamed; no
       repository write of any kind occurred (read-only file tools
       only); no delegation, no subagents, no background tasks, no
       Herdr operations, no settings changes.

     COMPLETION: complete.
