alphacouncil-agent 1.0.4 → 1.0.13

package/skills/alphacouncil-agent/SKILL.md

@@ -27,8 +27,9 @@ after Stage 0.
   explicitly adds optional analyst seats, every added seat is equally mandatory. A mandatory
   evidence failure is a fail-fast barrier: persist the failure and final diagnostic artifacts,
   skip masters/debate/PM model calls, and terminate `incomplete`; do not synthesize around it.
-  Headless full has a hard 1800000 ms queue-to-terminal-persistence ceiling. It fails closed
-  at expiry and does not promise all-seat success when external services deteriorate.
+  Headless full runs at one of three depth tiers set by `council_pace`: `fast` 900000 ms,
+  `normal` (default) 1800000 ms, `slow` 3600000 ms, queue-to-terminal-persistence. It fails
+  closed at expiry and does not promise all-seat success when external services deteriorate.
 - **Quick council (`quick_v1`)**: only the plugin-managed headless `analyze_symbol` path may
   execute it. It runs the four fixed evidence roles in parallel, 1-4 selected methods in
   parallel, one parallel bull/bear statement, and one short PM inside the hard 600000 ms
@@ -70,16 +71,44 @@ can do. It does **not** create a selection session, prove that the individual ca
 displayed, or issue a receipt, so it never substitutes for the steps below.
 
 The analysts have a sensible default -- the eight-seat fan-out -- and asking about them every
-time is a question with an obvious answer. Master selection is the one configuration decision:
+time is a question with an obvious answer. The gate takes two decisions in ONE interaction:
+which methods sit on the bench, and which depth tier the run uses.
+
+**Ask the tier; never make the user type `fast` or `slow`.** `begin_council_selection` returns
+`pace_options`, one row per tier with `expected_minutes`, `hard_ceiling_minutes` and what the
+extra time buys. Show it as a short menu ABOVE the seat catalog, because it is a three-option
+question and the catalog is long. Publish BOTH numbers: the ceiling on its own reads as the
+estimate, and then every fast run looks like it takes fifteen minutes. Render it like this, in
+the run's language:
+
+```
+本次分析要跑多深？（默认 2）
+  1. 快速   预计 ~12 分钟（上限 15）  每证据席 3.5 分钟，每轮辩论每侧 90 秒
+  2. 标准   预计 ~20 分钟（上限 30）  每证据席 6 分钟，每轮辩论每侧 150 秒   ← 默认
+  3. 深入   预计 ~44 分钟（上限 60）  每证据席 12 分钟，每轮辩论每侧 360 秒
+三档都是完整评议：同样 8 个证据席、同样三轮辩论、同样 PM。只有每席能想多久不同。
+```
+
+Pass the answer as `council_pace` to `confirm_master_selection`; it binds into the receipt, so an
+execution call may repeat it but never change it. If the request already said a speed, pass it as
+`council_pace` to `begin_council_selection`: that is a PREFILL exactly like a named master --
+highlight the row, still show the menu, still take the answer. No answer means `normal`.
+
+Quick returns an empty `pace_options` and rejects the field: it is a smaller contract, not a
+slower one. Say that plainly if a user asks for a fast quick run.
+
+Then take this run's method selection:
 
 1. Call `begin_council_selection` with `symbol`, the original `prompt`, inferred `language`,
    the calling `host`, and the intended `council_mode` (`full` by default). If the request
    explicitly names masters, resolve their stable IDs and pass them as
-   `preselected_master_ids`; preselection highlights only and never confirms.
-2. Display **every returned master individually in the returned order**. Preserve the stable
-   number and show `identity`, `method`, `best_for` and `maturity` for every row. A school
-   summary, preset or seat count does not satisfy this step.
-3. Take one submission. The universal fallback is a numbered text reply: one index in
+   `preselected_master_ids`; if it named a speed, pass `council_pace`. Both are prefills that
+   highlight only and never confirm.
+2. Display the `pace_options` menu with both numbers per tier, then **every returned master
+   individually in the returned order**. Preserve the stable number and show `identity`,
+   `method`, `best_for` and `maturity` for every row. A school summary, preset or seat count
+   does not satisfy this step.
+3. Take one submission covering both decisions. The universal fallback is a numbered text reply: one index in
    `1..N`, any comma/space-separated combination, ranges such as `1-4` or `1..4`, or stable
    IDs/names. Full also accepts `all`; quick requires 1-4 distinct methods and rejects `all`.
    - **Claude Code, Codex, OpenCode and Grok Build** may use a native multi-select when it can
@@ -94,7 +123,8 @@ time is a question with an obvious answer. Master selection is the one configura
    The obsolete rule **"Skip the question entirely"** is prohibited for council runs: a
    prefill reduces typing but never replaces this run's displayed catalog and receipt.
 5. Call `confirm_master_selection` with the exact `selection_id`, `catalog_hash`,
-   `display_ack: true`, and exactly one of:
+   `display_ack: true`, the answered `council_pace` (omit to accept `normal`), and exactly one
+   of:
    - `selected_master_ids: [...]` for a native multi-select;
    - `select_all: true` for all in full mode only;
    - `selection: "1,4-6"` (or another supported text selection) for the fallback.
@@ -117,18 +147,44 @@ Use this contract when full runs through headless `analyze_symbol`:
 - The eight mandatory evidence roles start together. Each has one bounded parse-only repair;
   repair converts malformed output and does not repeat web research.
 - After the evidence barrier, every selected physical v3 method executes its deterministic
-  policy and freezes its stance. Exactly one isolated voice worker is then launched for that
-  stable ID. It may explain the recorded policy result in the user's language, but cannot
-  change the stance, invent a typed fact or speak as the real named person. A missing voice
-  result remains visible and prevents a false complete bench.
+  policy and freezes its stance. Each seat that reached a stance then gets exactly one isolated
+  voice worker for that stable ID. It may explain the recorded policy result in the user's
+  language, but cannot change the stance, invent a typed fact or speak as the real named person.
+  A missing voice result remains visible and prevents a false complete bench.
+- A seat frozen `out_of_scope` is settled without a worker. Its deterministic statement names
+  the condition that closed its gate and states that an abstention is not a bearish vote, which
+  is all an out_of_scope seat is asked to say; `ALPHACOUNCIL_VOICE_ABSTAINING_SEATS=1` restores
+  a worker for every seat. Such a seat is still published with a readable statement, and its
+  `voice_status` is `deterministic_scope` rather than a claim that a worker ran.
+- Every condition id a seat cites is resolved back to the id its pack declares. The ids are
+  hashed before the policy runs so the decision layer cannot recognise the seat, but past the
+  freeze the seat is named in the report and in its own worker prompt, so a surviving
+  `anon_<hash>` only stopped a seat from telling a reader which condition decided it.
 - Round 1 Bull/Bear run together; after both pass, Round 2 runs together; after both pass,
   Round 3 runs together with exact saved-question bindings. The PM starts after both Round-3
   sides pass.
-- The hard ceiling is 1800000 ms from durable queueing through terminal artifact persistence,
-  including queueing, retries, all workers and deterministic finalization. A caller or
-  environment may lower it, never raise it. At expiry persist `incomplete` and name every
+- The ceiling is the selected tier's total from durable queueing through terminal artifact
+  persistence, including queueing, retries, all workers and deterministic finalization: `fast`
+  900000 ms, `normal` 1800000 ms, `slow` 3600000 ms. A caller or environment may lower the
+  selected tier's budget, never raise it. At expiry persist `incomplete` and name every
   timed-out/failed/skipped role. The deadline guarantees a terminal saved run, not successful
   completion under provider/search/data degradation.
+- The tier moves every per-stage cap with the total, because those caps are what bound each
+  worker: evidence 3.5/6/12 minutes per seat, method 1/2/4 minutes per seat, debate 90/150/360
+  seconds per round, PM 2/3/8 minutes. Raising `total_timeout_ms` alone buys idle time rather
+  than depth; the tier is what buys depth. All three tiers are `full_v2` — same eight evidence
+  seats, same three rounds, same PM — so a tier changes how long each seat may think, never what
+  the council is. Quick rejects `council_pace`.
+- The tier also shapes each worker's output, and this is the part that makes `fast` fast rather
+  than merely short of time. A cap alone is a timeout, and the same prompt with a shorter fuse
+  produces a packet the worker could not finish. Since an LLM call's wall clock is dominated by
+  the tokens it generates, `fast` asks for the same information in less prose: at most six
+  arguments, one to two sentences each, citing a source ID instead of re-quoting the evidence,
+  no opponent recap and no methodology preamble — while every figure, every scoped source ID and
+  every required report section stays mandatory, and price levels and invalidation conditions may
+  not be compressed. Dropping an argument is acceptable at `fast`; dropping a source ID or
+  filling a number from memory never is. `slow` instead asks for the derivation written out step
+  by step with explicit falsification conditions. `normal` adds nothing.
 - The concise handoff lists every selected stable master ID, frozen stance and voice-worker
   explanation/status; all eight analyst task IDs, statuses and summaries; and a system-owned
   price snapshot with currency/time/source or an explicit unavailable-data gap.
@@ -265,8 +321,11 @@ Use MCP only when the user explicitly accepts background/headless execution, wan
    `wait_for_completion=false` when the user wants a long/short or portfolio decision saved
    under `~/.alphacouncil-agent/runs/`. This returns a small durable accepted response with
    `run_id`, `status_json`, and `events_jsonl`; acceptance does not mean the report is done.
-   For quick, do not pass task overrides and do not request `synthesis=false`. For full, the
-   1800000 ms global maximum and parallel/barrier topology in the full contract above apply.
+   For quick, do not pass task overrides and do not request `synthesis=false`. For full, pass
+   `council_pace` when the user asked for a speed (fast 15 min / normal 30 / slow 60); the
+   selected tier's global maximum and the parallel/barrier topology in the full contract above
+   apply. `total_timeout_ms` above the tier's total is rejected and names the tier that allows
+   it.
 3. Poll `read_run(run_id)` at a bounded interval until `status.status` is terminal:
    `complete`, `degraded`, `incomplete`, `needs_verification`, `needs_revision`, or `failed`.
    Surface meaningful phase changes, not every unchanged poll. Poll the same `run_id`; never
@@ -320,6 +379,8 @@ Detect the user's language from their request and propagate it to EVERY subagent
 ### Stage 0 — Plan (envelope only)
 Call `plan_visible_run` with `symbol`, `prompt` (original user request), `as_of`, inferred `language`, `council_mode: "full"`, and the Stage 0 `selection_receipt`. It returns `run_id`, the planned evidence agent specs, the selected master specs, the 3 debate agent specs, and artifact paths. This is planning only (SKILL step 9); do not treat it as execution. The tool rejects `council_mode: "quick"`.
 
+Every planned prompt is written to `<run>/prompts/` and each agent spec carries `prompt_file`. Check `prompts_inline`: when it is `false` the prompt bodies were deliberately left out of the result, because returning them together would exceed what a host accepts, and you must `Read` each `prompt_file` instead of the inline field. What drives that size is the grounding each prompt embeds rather than the seat count, so a run with a full macro series crosses the budget where a sparse one does not. A truncated or rejected plan result is never a reason to write prompts from memory.
+
 ### Stage 1 — Evidence fan-out (one turn, isolated context)
 In a SINGLE assistant turn, emit one `Task` (subagent_type: general-purpose) call for every evidence role returned by the plan. The default eight are `market_data`, `earnings_deep_dive`, `forward_expectations`, `quant_factor`, `valuation_long_short`, `news_industry_management`, `insider_sec`, and `ib_event_analysis`. Each subagent:
 - May use ONLY `WebSearch` + `WebFetch`. It must NOT call `@alphacouncil-agent`, `collect_evidence`, `analyze_symbol`, or `read_run` (leaf-worker rule, Boundaries).
@@ -353,7 +414,7 @@ Run the documented rounds, each as a parallel fan-out of `bull_researcher` + `be
 Persist each round via `record_visible_decision(run_id, role, packet)` so `all_agents.md` accumulates the full trace. DISPUTED/UNVERIFIABLE claims may appear in a thesis only with an explicit caveat.
 
 ### Stage 4 — Verdict + synthesize
-Run one `portfolio_manager` `Task` fed the verified evidence plus all three debate rounds. Record it via `record_visible_decision(run_id, 'portfolio_manager', packet)`, which writes `decision.json` + `final_report.md` and marks the run complete. A successful PM response has `handoff_contract=inline_user_response_v1` and returns `user_response_markdown`; use that Markdown as the final user-facing response body instead of replacing it with a shorter recap. Its last section is the exact selected-seat count and one readable method statement per seat, including deterministic `out_of_scope` statements. Then link the complete report and audit artifacts in an appendix. The saved full report still contains the Analyst Work Log, Bull/Bear Debate record, verification ledger, all mandated sections, data gaps, horizons and `<task>:<source_id>` source table.
+Run one `portfolio_manager` `Task` fed the verified evidence plus all three debate rounds. Its packet MUST carry `report_markdown` as the complete report body with every authored contract section; the tool rejects a packet that does not, before taking the idempotency lock, and the rejection lists each missing section with the heading to use. Do not send a PM packet without it and expect the report to be assembled from the summary. The master bench and any instrument-structure section are system-appended and are never asked of you. Record it via `record_visible_decision(run_id, 'portfolio_manager', packet)`, which writes `decision.json` + `final_report.md` and marks the run complete. A successful PM response has `handoff_contract=inline_user_response_v1` and returns `user_response_markdown`; use that Markdown as the final user-facing response body instead of replacing it with a shorter recap. Its last section is the exact selected-seat count and one readable method statement per seat, including deterministic `out_of_scope` statements. Then link the complete report and audit artifacts in an appendix. The saved full report still contains the Analyst Work Log, Bull/Bear Debate record, verification ledger, all mandated sections, data gaps, horizons and `<task>:<source_id>` source table.
 
 Honest limits: Task fan-out is best-effort, not a guaranteed workflow engine; enforce the barrier by polling artifacts, not by assuming. WebSearch/WebFetch is the only evidence channel (no financial API), so some numeric claims stay "narratively corroborated, not vendor-verified". This is the same auditable contract as the other paths — a stronger runner, not a different audit story.