REQUEST_TOKEN: sol-opencode214-plan-opencode-20261023
     EFFECTIVE_MODEL: Requested by plan roster: opencode/deepseek-v4.
     1-flash-free, max, Plan. Configured: opencode/deepseek-v4.1-
     flash-free, Plan mode (harness); max variant not independently
     readable. Evidence: harness model line and Plan-mode reminder.
     Upstream identity unverified.
     VERDICT: changes_required

     BLOCKERS:

     - Plan line 56 vs line 65 — the usage grammar \d+(?:\.\d+)?[KM]?
       \(\d{1,3}%\) accepts 6.7K (300%), yet the negative matrix
       requires that exact string to remain Idle. Expected: freeze the
       intended percentage range (for example (?:100|[1-9]?\d)%) or
       remove the 300% negative and state that percentage semantics
       are not validated. Evidence: line 56; line 65 malformed-usage
       list; acceptance line 74.
     - Plan lines 50 and 66 — the frozen rule is a schematic, not a
       literal regex. ( <account> / <provider> | OpenCode | <usage> )
       leaves grouping, alternation precedence, incidental spaces,
       anchor scope and the exact old field classes unspecified; "the
       extended regex above" cannot be implemented without inference.
       Expected: include the exact line_regex string with non-
       capturing groups, ^/$, no decorative spaces, and the exact
       account/provider branch. Evidence: line 50; line 66.
     - Plan lines 57 and 65 — optional spinner is claimed but only
       old-shape and decorated new-shape positives are listed. No
       undecorated OpenCode or usage-token positive is named; a
       regression making the spinner mandatory in the new arms would
       pass the stated matrix. Expected: add no-spinner positives for
       both new suffixes. Evidence: line 57 "only one optional spinner
       cell"; line 65 positives.
     - Plan line 65 — negatives are not pinned per alternation arm.
       Quoted/prefixed/embedded rows are not explicitly assigned to
       the usage arm, and the only trailing-suffix row is OpenCode (
       copied); no quoted/prefixed/trailing usage-token footer or new-
       shape incomplete-final-field row is named. Expected: add
       explicit usage-arm quote/prefix/trailing-suffix negatives and
       new-shape incomplete rows (missing/empty final field, empty
       mode/model). Evidence: line 65; acceptance line 74.

     SUGGESTIONS:

     - Add a lowercase opencode (and perhaps OPENCODE) negative to
       prove the exact-case token.
     - State that K/M are uppercase-only and percentages are integer-
       only; add a lowercase usage negative if that is intended.
     - Disambiguate the 🯦/🯧 esc stop shorthand in lines 26–27; it
       reads as one line with two spinner cells, while the capture has
       separate frames and the rule permits one.
     - Pin the stale-cache fixture to the accepted 2026.10.22.1 (or
       preserved 2026.06.10.1) and name the 29-case matrix file so
       "older cached manifest" is reproducible.
     - Say "two-commit freeze (plan, then reference)" rather than "two
       plan-only commits" to match the OMP loop.

     FILES_AT_RISK: Inside the fresh temporary source copy only: src/
     detect/manifests/opencode.toml, distribution/agent-detection/
     opencode.toml, src/detect/manifest/tests.rs, docs/next/website/
     src/content/docs/agents.mdx. No skill_lens/ product path changes.
     CONCEDED: Offline proof cannot distinguish a byte-identical
     bottom line; that limit is acceptable and documented. The non-
     captured usage variants are grammar probes, not live evidence. No
     live trial, deployment or installation is required. The plan SHA-
     256 was not recomputed under the native-file-only/no-shell
     constraint. The old Default / OpenCode positive, blocked priority
     300, bottom-line anchor, version bump, both manifests, real-
     loader/CLI/stale-cache outline and offline-only boundaries are
     adequately covered.
     COMPLETION: complete / no delegates / no background work
