1 Commits

Author	SHA1	Message	Date
Edward Cheng-I Wu	48bde6a0b1	feat(tiering): #517 model tiering — judgment/execution split, two opt-in directions, default untouched (#520) ... * feat(tiering): #517 model tiering — judgment/execution split, two opt-in directions, default untouched One tiering mechanism per the frozen #517 design (Lance Martin, "Cost effective harnesses with Fable", 2026-07-10): - Default (ARS_MODEL_TIERING unset): byte-equivalent — every agent stays model: inherit. Same opt-in philosophy as terminal_policies. - economy (frontier session): 13 execution-type agents dispatch one tier below the session model, floor Opus-class, never Sonnet; draft_writer flagged as the highest-savings / most quality-sensitive downgrade. - quality-boost (below-frontier session): judgment-type agents at the Stage 2.5/4.5 gates + final-review surfaces step up to the frontier tier; nothing is ever downgraded. - Both directions carry explicit no-op announcements; unknown values warn once and behave as unset. Tiers are relative positions, never hard-pinned model ids (v3.7.0 opus-floor lesson). Classification: 39 agents (26 judgment / 13 execution; issue header's 25/12 arithmetic corrected, membership unchanged) in scripts/model_tiering_manifest.json + the canonical table in shared/model_tiering.md, pinned to each other and to the *_agent.md files on disk by scripts/check_model_tiering.py (11 mutation tests; wired into spec-consistency.yml + local pytest manifest, 60 -> 61). Consumers: compact "## Model Tiering (#517, optional)" block in the four SKILL.md files; SETUP en/zh-TW env-table rows; prompt-caching guidance (same-stage calls reuse the same worker) in the canonical doc. No agent-file edits (sha256-locked bibliography_agent.md untouched), no schema change, no hook. Spec: docs/design/2026-07-12-517-model-tiering-spec.md All CI lints + 61-entry pytest manifest green locally. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FczvB1WJjfumdJTg9YyL3X * fix(tiering): #517 round-1 codex fixes — dispatch-as-subagent mechanism, exact-set lint, caching scope, wording unification Addresses the round-1 codex review (2 P1 + 6 P2; 2 no-change with reasons): - P1 routing reality: canonical doc now states the mechanism honestly — a tier is selectable only at subagent dispatch; when a direction applies to an inline role, the session dispatches it as a subagent pinned to the target tier (that IS the mechanism); impossible-dispatch falls open to inline-on-session-model with a one-line announcement. PERFORMANCE.md en/zh "no separate model routing layer" reconciled. - P2 caching-vs-default: same-worker guidance scoped to active directions; unset default byte-equivalent incl. dispatch shapes. - P2 semantics drift: quality-boost = jump TO the frontier (not one increment); warn-once wording unified; SETUP rows name the final-review surfaces. - P2 lint subset-only + hardcoded dirs: exact per-(tier,skill) token-set comparison (missing/extra/duplicate tokens, per-row counts, duplicate rows) + repo-wide stray sweep; tests 11 -> 15. - P2 stale verification record: 60 -> 61 manifest entries in the spec. - No-change (documented): the PR-body session link is the repo's established harness attribution convention (48 occurrences in recent main history); Lance Martin/Anthropic is published-source author attribution consistent with the repo's citation norm. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FczvB1WJjfumdJTg9YyL3X * fix(tiering): #517 round-2 codex fixes — dispatch-time tier resolver, claim-ref audit surface, Stage 3' roster, PERFORMANCE cost qualification - P1 tier resolver: new "Resolving a tier at dispatch time" section — the no-hard-pinning rule governs repo files; the dispatch call resolves the relative target from the runtime's own model information (alias or current-generation id, ephemeral only); unresolvable -> announced no-op. - P1 claim_ref_alignment_audit: quality-boost surface corrected to its real dispatch slot (opt-in Stage 4->5, ARS_CLAIM_AUDIT=1), removed from the 2.5/4.5 gate list; SKILL blocks + SETUP rows updated. - P2 Stage 3' roster: canonical text now matches ARCHITECTURE (narrow re-review team; judgment roles there = eic + editorial_synthesizer); caching example rewritten accordingly. - P2 PERFORMANCE en/zh: plugin-agent cost-unchanged claim qualified with the unset condition + economy behavior for plugin-exposed execution agents (report_compiler). - P2 PR body refreshed (dispatch-as-subagent wording, 15 tests). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FczvB1WJjfumdJTg9YyL3X * fix(tiering): #517 round-3 P2 residuals — caching example covers the full 3' narrow team; PR body audit surface - Caching example names the authoritative Stage 3' re-review roster (field_analyst + eic + editorial_synthesizer) and states the reuse rule is tier-independent — field_analyst is the economy-affected role. - PR body quality-boost bullet gains the opt-in Stage 4→5 claim-ref audit surface (was omitted after the round-2 canonical correction). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FczvB1WJjfumdJTg9YyL3X --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>	2026-07-12 03:51:12 +08:00