From 5367b29188a3e84ef4a3da22685834fd10f5d28f Mon Sep 17 00:00:00 2001 From: bastien Date: Mon, 28 Sep 2026 18:17:34 +0200 Subject: [PATCH] =?UTF-8?q?docs(spec):=20effort=20tiering=20design=20?= =?UTF-8?q?=E2=80=94=20session=20high,=20agent=20pins,=20skill=20effort,?= =?UTF-8?q?=20phase=20shifts,=20max=20at=20loop=20caps?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../specs/2026-09-28-effort-tiering-design.md | 241 ++++++++++++++++++ 1 file changed, 241 insertions(+) create mode 100644 docs/superpowers/specs/2026-09-28-effort-tiering-design.md diff --git a/docs/superpowers/specs/2026-09-28-effort-tiering-design.md b/docs/superpowers/specs/2026-09-28-effort-tiering-design.md new file mode 100644 index 0000000..05fab60 --- /dev/null +++ b/docs/superpowers/specs/2026-09-28-effort-tiering-design.md @@ -0,0 +1,241 @@ +# Effort tiering — design + +Date: 2026-09-28 · Branch: `feature/effort-tiering` · Status: draft for review + +## 1. Intent + +Adapt the reasoning effort along a development run, not hold the whole +session at `xhigh`. The user's five-rung scale is the contract: + +| Rung | User definition | Examples | +|---|---|---| +| low | fix a line, rename a file, run a script | journal, commit, release bookkeeping | +| medium | day-to-day work | implement a closed plan, orchestrate between dispatches | +| high | a refactor, a bug that resists | investigation, diagnosis, contract drafting | +| xhigh | architecture, audit before validation | brainstorm, plan, challenge synthesis, gates | +| max | a stuck error, an error that cannot be recovered, or judged need | loop caps, error recovery | + +Automatic wherever the harness allows it. Where it does not, the user gets a +one-keystroke lever, never a silent default. + +Effort is a second axis on the BDR-077 routing table: BDR-077 fixed WHICH +MODEL runs each role and forbade inherit; this design fixes HOW HARD it +thinks, with the same no-inherit principle. + +## 2. What the harness allows (verified on Claude Code 2.1.283, 2026-09-28) + +Sources: code.claude.com/docs (model-config, skills, sub-agents, hooks), +the CHANGELOG (2.1.120, 2.1.149, 2.1.267, 2.1.280) and live probes in this +repo. + +| Mechanism | Verified behaviour | Evidence | +|---|---|---| +| Session level | Resolution order: `CLAUDE_CODE_EFFORT_LEVEL` env > `--effort` / `/effort` > settings (`modelSettings` per model, else top-level `effortLevel`) > model default (`high` on Fable 5.1). `max` is session-only, never persisted. `/effort auto` clears the per-model saved level only; a top-level `effortLevel` still applies. | docs | +| Subagent frontmatter `effort:` | Applied to the subagent. Absent → **inherits the session level**. | built-in on sonnet printed `xhigh`; impeccable agent pinned `medium` printed `medium` | +| Skill frontmatter `effort:`, user-typed `/skill` | Applied for the **rest of the turn**, AskUserQuestion included. | headless `/effort-probe-low`: every request at `low` | +| Skill frontmatter `effort:`, loaded by Claude through the Skill tool, **interactive** session | Applied for the rest of the turn. Last loaded skill wins, up and down. | this session: `xhigh` → probe max → `$CLAUDE_EFFORT=max`, request records `effort=max` → probe xhigh → back to `xhigh` | +| Same, **headless** (`-p`) | **Not applied** (neither `effort:` nor `model:`). | three `-p` runs, transcript effort unchanged | +| Prompt cache on a mid-turn shift | **Preserved** on Fable 5.1: first request at max read 206,996 cached tokens, wrote 1,164. | this session | +| Agent tool call site | No `effort` parameter (only `model`). One agent file = one effort. | tool schema | +| Hooks | Read `$CLAUDE_EFFORT` / `effort.level`; **cannot change** the level. | docs | +| `ultrathink` keyword | In-context nudge only; the effort sent to the API is unchanged. | docs | +| Env var | `CLAUDE_CODE_EFFORT_LEVEL` beats every frontmatter override. Unset on this machine. | docs + `env` | + +## 3. What the numbers say (6 days of local transcripts, all projects, 10,955 requests) + +Weights relative to input price: output ×5, cache read ×0.1, cache write ×1.25. + +| Item | Share | +|---|---| +| Cache reads (context re-read per request) | 53 % of weighted spend | +| All output tokens | 16 % | +| of which thinking | 8 % | +| Thinking located in the main loop | 97 % of thinking | +| Mean thinking per request: Fable main loop / sonnet subagent at xhigh | 1,430 / 26 tokens | +| Mean cached context per main-loop request | ~320 k tokens | + +Consequences. Executors barely think even at xhigh: pinning them is about +explicitness and future models (Opus 5.5 "thinks more per turn at a given +level"), not savings today. The direct lever of effort is single-digit +percent; the indirect lever (fewer steps at lower effort → fewer requests → +fewer cache reads) is unmeasured and gets an A/B in §9. The dominant cost is +main-loop context size, out of scope here (see `/capitalize`, `/clear`). + +## 4. Decisions + +### D1. Session default `high` +`settings.json` `effortLevel`: `xhigh` → `high`, explicit rather than +deleted: the statusline reads the key, and LRN-139 wants a visible value to +sweep at every model bump. Interactive chat outside a skill runs at the +model default; the user raises with `/effort xhigh` (session) or the new +`/effort-max` shifter (turn, see D4). `CLAUDE_CODE_EFFORT_LEVEL` must stay +unset (it would silence every override below); the session-start banner +warns if it is set. + +### D2. Agent pins (approach A) — repo-authored agents only + +| effort | Agents | +|---|---| +| low | hotfixer, release-executor, plugin-probe, validator-analyzer | +| medium | feater, bugfixer, code-cleaner, onboarder, scaffolder (was `high`; citer `skills/init-project/SKILL.md:98` updated) | +| high | refactorer, analyzer, commit-changer, doc-syncer, handover-doc-writer | +| xhigh | plan-challenger, plugin-advisor, verifier, security-auditor, seo-analyzer, geo-analyzer | +| none | interviewer, client-handover-writer (inline-load only, a pin would be inert and misleading, BDR-076 precedent); status-reporter (haiku, no effort support); `impeccable-*` (vendored) | + +Rules. One effort per agent file, so a mode-based agent (BDR-077) pins the +level of its **judgment** mode and its mechanical modes over-tier: the +fail-safe direction, and free on sonnet per §3. Built-ins (Explore, +general-purpose, Plan) cannot be pinned at the call site and inherit the +main loop's current level; Explore on Fable thinks ~1 token per request, +so no wrapper agent is created. Verifier and security-auditor sit at xhigh +by the user's own definition ("audit before validation"); on sonnet the +cost difference is nil. + +### D3. Skill frontmatter effort (approach B) — the run's entry level +Applies from the user's invocation for the rest of the turn. + +| effort | Skills | +|---|---| +| low | status, commit-change, release-candidate, doc, capitalize, close, reconcile, deploy, profile, plugin-check | +| medium | gitflow, prune-memory, find-docs | +| high | feat, hotfix, bugfix, refactor, web-validate, harden, seo, geo | +| xhigh | ship-feature, init-project, onboard, tour, audit-delta, analyze, code-clean, client-handover, spec, skillify, brainstorming, writing-plans | +| unlisted | session default, by design (gstack and plugin skills are external; graphify is machine-owned) | + +`brainstorming` and `writing-plans` are vendored superpowers skills: the +one-line patch drifts from upstream at each resync; a census lock (§7) +makes the loss loud. + +A skill loaded by Claude as a sub-step (feat → commit-change) also shifts +the level for the rest of the turn (interactive, §2), so orchestrators +re-assert their own level after any nested Skill call whose level differs +(D4 protocol). + +### D4. Phase shifts inside a run (approach C) +Five one-line skills, no body beyond a sentence, user-invocable: +`effort-low`, `effort-medium`, `effort-high`, `effort-xhigh`, `effort-max`. +Descriptions as pre-validated against the routing census (pairwise +similarity ≤ 0.03). Protocol in a shared include `lib/effort-shift.md`, +mirroring `lib/model-gate.md`: + +- A shift is a `Skill(effort-)` call on the main loop. Never inside a + dispatched agent (agents run on their pin). One tool round-trip, + cache-safe (§2). +- Orchestrator wiring, three points each: `effort-medium` when the plan is + closed and the dispatch phase starts; `effort-low` before the + capitalize / journal / doc-commit tail; `effort-max` at an escalation + point, then the skill's own level again once the diagnosis is produced. +- Re-assert the skill's own level after any nested `Skill(...)` call whose + frontmatter carries a different effort (D3): the nested level would + otherwise hold for the rest of the turn. +- **Escalation points (automatic max)**: verify-secure loop GATE 1 cap + (3 conformity rounds) and GATE 2 cap (3 security rounds), before the + human-escalation table is composed; ship-feature STEP 4b, so the + inline analyzer DEBUG read runs at max. Full conversation context is the + asset here; a fresh diagnoser agent was considered and dropped (YAGNI: + no context, one more agent, same effort). +- **Not automatic, by doctrine**: the challenge fail-safe (a mute + challenger is an infrastructure failure, not a reasoning problem) and + the "gone WRONG → STOP" rule (STOP precedes any further reasoning). Both + STOP messages name the level reached and suggest `/effort-max` for the + relaunch: a turn-scoped max the user gets by typing one command. +- **Turn reset**: a prose gate that ends the turn (model-gate STOP, loop + cap STOP, and the four prose gates found in bugfix, ship-feature ×2, + init-project) drops the resumed turn to the session level. The plan + audits each such gate: if the resumed phase is reflection, the resume + step re-asserts with `Skill(effort-xhigh)`; if it is dispatch or + orchestration, session `high` is adequate and nothing is added. +- **Headless limitation**: `-p`, `claude agents` and SDK sessions ignore + skill-level effort (§2); runs there stay at the session level. Documented + in the include, no mitigation. + +### D5. Visibility +`hooks/statusline.sh` shows `$CLAUDE_EFFORT` when set (the live level, +shifts included) and falls back to the settings key. `/tasks` already shows +each subagent's effort (2.1.243). + +## 5. Alternatives rejected + +- **Keep xhigh, pin executors only**: executors think ~26 tokens per + request; the burn is in the main loop (§3). +- **Escalation diagnoser agent (`model: fable`, `effort: max`)**: chosen + before the interactive probe proved C viable; dropped because the + main-loop shift keeps the full failure context and adds no agent. +- **Move reflection into `model: fable` skill-runner children with + `effort: xhigh`, session at medium**: loses conversation context and + interactivity (BDR-077 retention criteria), heavy re-architecture for a + lever C delivers in five one-line files. +- **Rewrite settings.json mid-run to shift effort**: global side effect on + every session, LRN-098 drift class, fights the harness. +- **`maxEffortLevel` cap on sonnet**: pins already bound each agent; a cap + would hide a mis-pin instead of failing it in the census. + +## 6. Files touched + +| Area | Change | +|---|---| +| `settings.json` | `effortLevel` → `high` (curated config: read the diff, LRN-098) | +| `agents/*.md` (20) | `effort:` line per D2; `skills/init-project/SKILL.md:98` citer | +| `skills/*/SKILL.md` (33) | `effort:` line per D3, including the two vendored superpowers skills | +| `skills/effort-{low,medium,high,xhigh,max}/SKILL.md` | new, frontmatter + one sentence | +| `lib/effort-shift.md` | new include: protocol, wiring points, escalation, turn reset, headless note | +| `lib/model-gate.md` §4 | one paragraph: effort is the second axis, pointer to the include | +| `lib/verify-secure-loop.md` | `Skill(effort-max)` before each cap's human-escalation table; STOP text names the level | +| `lib/challenge-plan.md` | STOP text names the level, suggests `/effort-max` | +| orchestrator SKILL.md (feat, hotfix, bugfix, ship-feature, init-project, onboard, tour, code-clean, seo, geo, harden, web-validate, client-handover, audit-delta) | include line + the three wiring points; ship-feature 4b max | +| `hooks/statusline.sh`, `hooks/session-start.sh` | live effort display; env-var warning | +| `lib/tests/effort-routing.test.sh` | new census suite (§7) | +| `CHANGELOG.md`, `.claude/memory/*` | release note; BDR + LRN + EVAL + journal (§8) | + +## 7. Tests and census (`make test`) + +New suite `lib/tests/effort-routing.test.sh`, `grep -qF` locks in the +`model-routing.test.sh` style, flip-tested first (BDR-100): + +1. Every repo-authored agent outside the "none" list has `effort: ` + in its first 10 frontmatter lines, level in the allowed set; the "none" + list has no `effort:`. +2. Tier locks per D2 (one `has` per agent). +3. Skill locks per D3 (one `has` per skill), including `brainstorming` and + `writing-plans` (the resync alarm). +4. The five shifter skills exist with the exact `name:` and `effort:`. +5. `lib/effort-shift.md` is included by every orchestrator in the §6 list; + `verify-secure-loop.md` and `ship-feature/SKILL.md` contain the + `Skill(effort-max)` lock. +6. `settings.json` `effortLevel` is `high`. +7. `skill-routing-census` stays green with the five new descriptions + (pre-validated). +8. `doctrine-citers` stays green: no new `CLAUDE.md "…"` citation; the + doctrine lives in `lib/`. + +Per-wave smoke, planted input, disk-verified (BDR-077 precedent): +W1 dispatch a pinned agent that echoes `$CLAUDE_EFFORT`; W2 invoke `/status` +and read `effort=low` in the transcript; W3 run a skill through a shift +and read the request sequence; W4 statusline shows the live level. + +## 8. Rollout + +Four waves on `feature/effort-tiering`, one commit each, smoke as merge +gate, human signal for `gitflow finish`: + +- W1 settings + agent pins + test suite + model-gate paragraph. +- W2 skill frontmatter (D3) + superpowers patch. +- W3 shifter skills + `lib/effort-shift.md` + orchestrator wiring + + escalation points + turn-reset audit. +- W4 statusline + banner warning + CHANGELOG + registries. + +Registries: BDR (effort tiering, this spec's decisions and rejected +alternatives), LRN (skill effort applies on user invocation and on +interactive Skill-tool loads, not in `-p`; shifts are cache-safe), EVAL +(the §3 measurement and its method), journal line. + +## 9. Measurement after rollout + +A/B on a repeatable skill run (`/reconcile` on this repo, session `high` +vs `xhigh`): requests, output tokens, thinking tokens, wall time, from the +transcript. Records whether the indirect lever exists. Goes to EVAL. + +## 10. Out of scope + +Main-loop context size (the 53 %), gstack and plugin skills, the +`impeccable-*` agents, `graphify` (machine-owned), headless sessions.