REQUEST_TOKEN: sol-mini-trial-plan-opencode-20261023

     EFFECTIVE_MODEL: requested opencode/deepseek-v4.1-flash-free#max
     / Plan / read-only; configured opencode/deepseek-v4.1-flash-free
     / Plan per this session; max variant not independently read back;
     evidence: harness model/provider metadata and enforced native-
     read-only tools; upstream identity unverified.

     VERDICT: changes_required

     BLOCKERS:

     1. Acceptance line 62; Execution steps 4 and 7 (lines 47, 50) —
        the clock comparison is not operational. The plan records
        whether time.completed is seconds or milliseconds but sets no
        resolution or clock-offset guard. At second resolution, a wait
        returning up to one second before true completion can still
        compare as “not before” after truncation; equality and sub-
        resolution differences are undefined. Expected: record the
        clock source/domain and timestamp resolution; require
        wait_return >= completed + one resolution unit (or a measured
        offset), and treat any difference inside that band as
        inconclusive. Evidence: line 47 admits the API does not
        promise a unit; line 62’s “timing ambiguity” has no test.

     2. Acceptance line 62; Execution step 5 (line 48) — the original
        wait need not succeed. The pass list omits a successful
        settled agent prompt --wait result. A non-timeout error or
        unknown return after native completion could satisfy the
        timing, token and finish:stop checks. Expected: pass requires
        the wait to return normal idle/done (not blocked/error/
        timeout/unknown) after completion; any non-zero or uncertain
        result is failure/inconclusive. Evidence: line 51 lists
        timeout/stall only; line 62 has no result-status clause.

     3. Execution steps 3 and 7 (lines 46, 50) — no exactly-one-
        delivery proof. “Single empty native session” and “one prompt”
        are asserted, but post-prompt evidence only filters the last
        assistant message. It does not require zero user messages
        before submission, exactly one token-bearing user message, and
        no other turns. A duplicate Herdr submission or startup-
        created turn could still yield a complete matching answer.
        Expected: record the pre-prompt count; after, require exactly
        one token-bearing user message, one matching final assistant,
        and no other turns; otherwise fail/inconclusive. Evidence:
        line 38 asserts emptiness without a verification step; line 50
        checks only the last assistant.

     4. Execution steps 5–6 (lines 48–49) — concurrency of the one
        prompt wait is undefined. The busy observation must occur
        “while that same wait is outstanding,” but agent prompt ... --
        wait is described as a blocking command. The plan does not say
        how it is supervised/backgrounded, how its exact return time
        and exit status are captured, or how the observer runs before
        the turn ends. Sequential execution misses the busy snapshot
        by construction. Expected: specify a supervised background
        launch (PID/start/output/exit), then the single observer; no
        second submission. Evidence: line 48 uses --wait; line 49
        assumes concurrency; line 55 bounds durations only.

     5. Scope steps 4 and 6 (lines 34, 36) — preflight order is
        contradictory. Step 6 says model availability must be proved
        “before any Herdr build,” but step 4 builds and starts the
        corrected Herdr server. Literal order breaks the preflight
        rule and can build/start before discovering the model is
        absent or needs a credential. Expected: run the model
        preflight before step 2, or hard-gate steps 2–4 on it; stop
        before any build/server if unavailable. Evidence: line 36’s
        bold “before any Herdr build” versus line 34’s step 4.

     6. Reporting line 66 versus Acceptance line 62 — final critics
        are not an acceptance condition. Two fresh critics are
        required, but the pass list does not require both to return
        pass, and gives no rule for one changes_required/unverified
        beyond vague “two nonconvergent rounds” escalation. One pass
        plus one unverified could be reported as a completed pass.
        Expected: live pass only if both critics return substantive
        pass on the same frozen evidence; dissent or unverified
        blocks/escalates; tie each report to native completion and no
        active work. Evidence: line 66 says “assess,” not “both pass”;
        OMP review/dispatch rules require blockers resolved and
        substantive reports.

     SUGGESTIONS:

     - Add a tri-state table (pass/failure/inconclusive) matching
       Outcome line 13; acceptance lines 62–64 collapse the last two.
       Define, for example, early wait as failure and missed busy
       snapshot or clock ambiguity as inconclusive.
     - Before the prompt, verify the live evidence commands (agent
       wait --until working, agent explain, agent get, agent read)
       against installed 0.9.5 help and the idle isolated pane;
       current preflight covers only the OpenCode API.
     - Require a substantive ~500-word answer or a minimum length/
       content check, not only BEGIN/END markers.
     - Align line 68 with docs/process/omp/planning-loop.md §5: commit
       1 plan-only, commit 2 reference-only; “two plan-only commits”
       is ambiguous. Record the fresh token in the ledger.
     - Record a pre/post wall-clock bracket and monotonic readings
       around the wait; a clock step is inconclusive.
     - Verify no first-run/onboarding modal is present before
       submitting the prompt.

     FILES_AT_RISK:

     - docs/plans/2026-10-23-herdr-mini-isolated-trial.md
     - /tmp/skill-lens-herdr-mini-live-20261023-* (owned source/tools/
       homes/sockets/logs)
     - /Users/corvette/.config/herdr/herdr.sock, user default session,
       pane wG:p1
     - /Users/corvette/.local/bin/herdr; /Users/corvette/.opencode/
       bin/opencode
     - User shared OpenCode service at http://127.0.0.1:49374
     - docs/validation/omp-coordinator/2026-10-23-herdr-mini-isolated-
       trial/
     - Preserved modified docs/validation/omp-coordinator/2026-10-20/
       herdr-dispatch.md, untracked .path, untracked 2026-10-19 plan.
     - origin/main if evidence is committed/pushed.

     CONCEDED:

     - Debug build only, one free model (not the historical MiniMax
       alias), one live attempt, no install/release proof, no full-TUI
       trace/fix, no formal smoke, no general Herdr reliability claim;
       upstream served identities unverified. The plan correctly
       treats the accepted offline correction as input evidence and
       preserves the normal services/files.
     - I did not compute the expected plan SHA-256 with native read
       tools; the coordinator’s pre/post check owns it.
     - Process note: a broad read-only search surfaced preserved
       receipt text; no transcript content is used or quoted, and no
       credentials/auth files were opened.

     COMPLETION: complete / no delegates / no background work
