reliquary/plugins/spire/skills/implementation/references/models.md

Model notes

Observations about models in Spire work. They are dated and sourced, and they are not law. Use them to pick a default and to know what to watch for; the problem at hand decides. When a note is contradicted, add the newer observation with its date rather than silently dropping the old one, unless it is plainly withdrawn.

Context figures are guesstimates from Claude Code experience. Agents cannot see their own token count, and orchestrators cannot yet look into a running subagent's usage.

Sources: dirigent docs/workpad/reports/2026-10-01-lean-process-report.md (+ 2026-10-05 addendum); spacecraft docs/workpad/reports/2026-09-27-13-strategy-mvp-retrospective/report.md, docs/workpad/reports/2026-09-29-13-orchestration-overhead-attribution.md and docs/workpad/brainstorm/the_assistants_mindset.md; reliquary journal docs/workpad/journals/2026-10-05-21-athanor.md (user direction).

Defaults (2026-10-05)

Opus 5.5 for the Athanor, planning, attacking plans and phase review (user direction). Owner models for blocks come from the build plan; a common split is Opus 5.5 for hotspot blocks (state, concurrency, formats, read-path precedence) and Sonnet for mechanical ones (field sweeps, wiring, docs). Housekeeping (test runs, commits) on Haiku.

Opus 5.5 — claude-opus-5-5

• Catches more, decides boldly. Can replace frequent small reviews, especially in Rust. (user, 2026-10-05)
• Can take the user's word as holy, or reverse 100% when corrected. Relay user remarks with their weight: a passing thought is not a rule. (user, 2026-10-05)
• Plans are thorough and long. A 5.5 plan traced a root cause in code that a 4.6 plan on the same scope missed (dirigent, 2026-10-01). 5.5 plans ran about 11× larger than earlier plans and drove scope (spacecraft, 2026-09-27). The attack step trims scope. Dirigent reports properly sized plans since (user, 2026-10-05).
• An independent 5.5 review found 6 Important issues in about 20 min where an in-furnace review had passed; one 5.5 fix pass cleared them in 33 min (dirigent, 2026-10-01).
• Context: comfortable to about 600k (guesstimate, user, 2026-10-05).

Opus 4.6 — claude-opus-4-6

• Fast and capable. Sometimes better than 5.5 at interpreting what the user means; useful as an "understanding the human" interpreter. To be replaced eventually. (user, 2026-10-05)
• A 4.6 plan misidentified a root cause and would have broken archive compatibility; one plan, one comparison (dirigent, 2026-10-01).
• The dirigent Athanor end review that passed with zero findings ran on 4.6 (user's attribution, 2026-10-05; the report recorded "model unknown").
• Withdrawn 2026-10-05: "prefer Opus 4.6 for implementation" and "Opus 5 tends to hallucinate" (former Igor model-tier notes, 2026-09). The 5.5 evidence above contradicts them.

Fable

• Better still at interpreting intent (user, 2026-10-05). Used for two design reviews in spacecraft (Fable 5.1, 2026-09-27). Full id not recorded here.
• Not dispatched as a subagent unless the user asks (Igor rule).

Sonnet

Full id not recorded here; spacecraft notes mention "Sonnet 5".

• Dirigent Sonnet-first-draft experiment (2026-09-28 to 2026-10-01): the mechanical blocks were clean; all 6 Important findings sat in complex blocks. About 4 of the 6 were plan gaps that any implementer following the plan would have carried out. The user ended the experiment and moved complex work to 5.5. Read this as "the plan mattered as much as the model", not as "Sonnet cannot build".
• Clean on run-and-record and checklist work. Wrote tests that could not fail; could not finish an open-ended performance task (spacecraft, 2026-09-29).
• Reached 500k–750k tokens on long complex tasks (dirigent memory, 2026-10-01).

Haiku

• Used for runs, compiles, commits and mechanical appliers. In one spacecraft run it reported "no deviations" while edits had landed in the wrong place (spacecraft, 2026-09-29).

Context and block size

• Both extremes failed. Many tiny fresh implementers relearn the code each time. One implementer carried across tasks reached about 960k tokens and drifted (spacecraft, 2026-09-29; model probably 5.5 through the alias).
• Measured: dirigent blocks of 11–36 min of agent work used 186k–344k tokens each (Sonnet, 2026-10-01). None ran out.
• Guesstimates (unmeasured): a trunk read at intake of about 100k–250k; leave room for fallout and the fix pass rather than running close to the comfortable limit.
• With gremlins (2026-10-16), the trunk read is the base's training, not the Athanor's context. Measured: the Phase 0 large base was 182k tokens (P0.2) and the smoke base 23.7k; the Athanor reads plans and reports, not code (preflight report; Phase 3 contract A13).

Future work: orchestrator introspection of subagent token use, and signals into running work (dirigent is working on this). Until then, the Athanor records what usage it sees in completion results and judges by block count and size.

• Update 2026-10-16: gremlins (headless sessions) report per-call usage through usage (tokens by cache column, models used, priced cost). Agent-tool subagents still do not.

Notes added 2026-10-16

• Plan mode (--permission-mode plan) silently upgrades a Haiku session to Sonnet; init.model still shows Haiku, so read modelUsage for what ran. The docs only imply it (research F2a, F2b).
• Aliases map to model versions per CLI release: haiku is Haiku 4.5 on claude 2.1.283 and Haiku 5.5 on 2.1.299 (research F2c). Pin a full id where the version matters.
• Opus 5.5 may refuse a bare "run this blocking command" brief as pointless; say why the command is wanted (observed in the Phase 0 trunk notes, preflight report, 2026-10-16).