# Runtime Budget Policy

> Shared policy for keeping AK-Threads-Booster usable on low-allowance agents.

Version: 2.0.0

---

## Purpose

Daily runs should use compact, source-linked summaries by default. The tracker and deep knowledge files remain the source of truth, but they are not the default runtime surface.

## Default Modes

Read `threads_booster_config.json` if present. If a key is absent, use:

```json
{
  "runtime": {
    "token_mode": "ask",
    "depth": "standard",
    "compiled_memory": "prefer"
  },
  "analyze": {
    "output_mode": "brief"
  }
}
```

## Token Mode Choice

If `runtime.token_mode` is absent or `"ask"` and the run is interactive, ask before doing any heavy reading:

```text
這次要用哪種模式？

1. 低 token 版（建議日常使用）
   優點：比較快、省用量，適合日常檢查、一般選題、普通草稿。
   缺點：主要讀 compiled memory + quick cards，細節比較少；遇到微妙風格或演算法邊界時，可能需要再切高 token 深挖。

2. 高 token 版
   優點：讀更多歷史資料與完整知識庫，分析更細，適合重要貼文、深度品牌聲音校準、重大策略決策。
   缺點：比較慢，也更耗 Agent 用量。

你可以回「低 token」、「高 token」、「這次低 token」、「這次高 token」。
```

Map the choice:

| User choice | Runtime mapping |
|---|---|
| low token / 低 token | `runtime.depth = "standard"`, `runtime.compiled_memory = "prefer"`, `analyze.output_mode = "brief"` |
| high token / 高 token | `runtime.depth = "deep"`, `runtime.compiled_memory = "off"` for the current run unless the skill only needs compiled memory as an index, `analyze.output_mode = "full"` |

If the user says "always low token" or "always high token", persist `runtime.token_mode` only from a write-capable config owner. If the current skill cannot write config, apply the choice for the current run and tell the user how to make it permanent.

If the run is headless or cannot ask, default to low token mapping. This protects low-allowance agents from surprise cost.

## `runtime.depth`

| Value | Behavior |
|---|---|
| `lite` | Read compiled memory + quick cards only. Use at most 3 comparable posts. Do not read full knowledge files unless a red line or factual uncertainty requires it. |
| `standard` | Read compiled memory + quick cards. Use 3-5 comparable posts. Read targeted tracker excerpts only when provenance is needed. |
| `deep` | Full analysis mode. Read tracker, relevant companion files, and full knowledge files as needed. Use when the user asks for depth or when compact context is insufficient. |

## `runtime.compiled_memory`

| Value | Behavior |
|---|---|
| `prefer` | Use `compiled/` files first. Fall back to tracker when missing, stale, or insufficient. |
| `require_fresh` | Use compiled memory only if its `source_tracker_hash` matches the current tracker hash. Otherwise rebuild or stop and ask. |
| `off` | Ignore compiled memory and use source files directly. |

## `analyze.output_mode`

| Value | Behavior |
|---|---|
| `brief` | Output only red lines, decision summary, pointed changes, highest-upside comparisons, AI-tone density, and reference strength. |
| `standard` | Output the normal analysis sections, but keep each section compact. |
| `full` | Output the complete 11-section report from `skills/analyze/references/output-format.md`. |

## Compiled Memory Files

Expected location: `compiled/` in the working directory.

- `account_wiki.md` - compact account baseline and source-linked conclusions.
- `account_state.md` - three-axis diagnosis: algorithm state, audience psychology state, and anti-AI state.
- `personal_signal_memory.md` - personal account signals grouped by algorithm, psychology, and anti-AI texture.
- `next_move_queue.md` - algorithm-gated next moves, each tied to S targets and R risks.
- `post_feature_index.jsonl` - one feature/metrics row per post.
- `cluster_wiki.json` - semantic clusters, freshness, representative posts.
- `exemplar_bank.md` - a capped set of representative posts.
- `recent_window.md` - recent posts and repetition risk.
- `voice_fingerprint.md` / `voice_fingerprint.json` - deterministic voice-pattern cache for `/voice` and `/draft`, including engagement weighting, temporal shifts, cognitive-layer seeds, anti-voice candidates, and calibration anchors.

These files are derived views. They are never the source of truth.

## Freshness Rules

Compiled memory is fresh only when:

1. `source_tracker_hash` matches the current `threads_daily_tracker.json` hash.
2. `posts_count` matches the current tracker post count.
3. `generated_at` is present.

If any check fails:

- `lite`: say compiled memory is stale and switch to tracker-only fallback for the smallest needed excerpt.
- `standard`: read tracker enough to verify the relevant claim, then recommend rebuilding compiled memory.
- `deep`: use source files directly and recommend rebuilding compiled memory after the run.

## Source-Truth Rule

When compiled memory and tracker disagree, the tracker wins. Report the mismatch briefly and rebuild compiled memory before using that compiled claim again.

## reason + strip_when

| Component | Reason | Strip when |
|---|---|---|
| Runtime depth modes | Lets low-allowance agents avoid loading full history and deep knowledge on routine runs. | Agent pricing/context limits stop being a practical constraint, or telemetry shows users consistently choose `deep`. |
| Compiled memory preference | Historical comparison sets are mostly deterministic and can be precomputed with provenance. | Tracker moves to a queryable store with native retrieval and cheap filtered reads. |
| Brief analyze output | Output tokens are a meaningful part of total cost; many runs only need the decision layer. | Users consistently request full reports, or UI-level collapsible sections make output tokens cheap. |