decide: decision point specs

Constants live at the top of decide.py and decision_points.py; questions, options and rubrics in points.json. Tests: 06-optional/tests/test_decide.py.

D-state

State sent with each question: one intro sentence + compact sorted-key JSON of facts, cut at MAX_STATE_CHARS (6000 chars, ~1.5k tokens). Facts are numbers already in sessions.jsonl / analysis.json / impact.jsonl; no prompt text, paths or repo names except repo_label. Cost mix = part / cost.total per type (input, output, cache_read, cache_write), 3 decimals, 0 when total is 0. Cache hit = cache_read / (input + cache_read + cache_write), null when the denominator is 0 (same as analyze M-cache-hit-rate). Lists: top TOP_TOOLS tools by calls, top TOP_MODELS models by cost.

D-cache

state_sha = sha256 of JSON [state, question, options, kind] (sorted keys), first 16 hex. A (point, state_sha) present in decisions.jsonl is never asked again; vendor is not part of the key. Changing a rubric or any fact changes the sha and asks again. Duplicate items in one run are asked once.

D-pick

choice = top option if p(top) - p(second) >= MIN_LEAD (0.2), else "unsure". Ties broken by option name. Single option: lead = p. Example: {a:.5, b:.3, c:.2} -> a; {a:.45, b:.35, c:.2} -> unsure; {yes:.9, no:.1} -> yes.

D-outlier-explanation

One choice item per analysis.outliers[i]. Facts: metric id and definition, group, value, group median (baseline), robust z, group n, and the session facts (D-state) when session_id is in sessions.jsonl. Subject <metric>:<group>:<session_id|->. Options: points.json outlier_explanation. Result in analysis.decided.json as outliers[i].decision = {choice, probabilities, state_sha, vendor}. Missing session: asked with outlier facts only. D-outlier-model: facts also carry baseline (model when the group is <harness>/<model>, harness_fallback when the group is the whole harness; then also baseline_typical_model = most common main model in that harness and its price tier), and, with a session, main_model (most-cost model), main_model_list_usd_per_mtok {input, output} from schemas/prices.json and main_model_price_tier by output price: <= 1 budget, <= 15 standard, else premium, no price -> unpriced. Without these, an expensive model scored against a cheap fallback median was labelled agent_loop.

D-recommendation

One noul item per (candidate, harness): every analysis.recommendations entry (its evidence added to facts) plus every catalog candidate in analyze/metrics.json recommendations (single source; points.json holds only the question) for each harness that has sessions and whose id is not already listed. Facts: harness aggregates from sessions.jsonl (sessions, cost, cost mix, cache hit, turns, calls, calls per turn, median session cost, subagent share, model cost share, MCP error rate). harness null -> all sessions. Subject = the recommendation id (<candidate id>/<harness>, the same id analyze writes, so a candidate is asked once per harness). Applied: p_yes = p(yes); verdict suggest if p_yes >= SUGGEST_P (0.7), skip if <= SKIP_P (0.3), else unsure; not yet answered -> p_yes and verdict null. Catalog candidates are appended with that id and evidence cost_share.<type>. Items for the same harness share a state, so a batching vendor answers them in one request.

D-commit-link

One noul item per impact.jsonl record with link:"window", link_p:null, non-empty commits, and its session in sessions.jsonl. Facts: session harness, start, end, active minutes, calls, turns, edit tool calls (EDIT_TOOLS in decision_points.py: file-editing tools only, shown to the model as evidence of edits; unlike impact WRITE_TOOLS, which adds shell tools because L-assign only needs "could have run git commit"), and per commit ts, files, added, deleted, agent_trailer, on_default_branch. Subject = L-subject in impact/METHOD.md. Runs whenever such records exist (impact needs it), regardless of decide.points. Impact reads probabilities.yes from decisions.jsonl (point commit_link, matching subject) as link_p and drops it below its MIN_P. Ledger tag impact.

Contract impact<->decide (commit_link): kept in one place, impact/METHOD.md L-subject.

D-intent

Reserved. Needs prompt text, which v002 events never carry; listing intent in decide.points stops the run.

D-apply

analysis.decided.json = deep copy of analysis.json with D-outlier-explanation and D-recommendation applied, validated against the analysis schema. analysis.json is never written. Each run asks at most MAX_CALLS questions (--max-calls); the rest stay unanswered until the next run. The first vendor error stops asking; answers so far are kept.