2026-03-23 12:47:32 +00:00
|
|
|
<!-- last-reviewed: eb7e24cb1df028c6061f47ddfdf9b4ebec33e1cf -->
|
2026-03-21 12:44:23 +00:00
|
|
|
# Predictor Agent
|
|
|
|
|
|
2026-03-23 13:56:59 +00:00
|
|
|
**Role**: Abstract adversary (the "goblin"). Runs a 2-step formula
|
|
|
|
|
(preflight → find-weakness-and-act) via interactive tmux Claude session
|
|
|
|
|
(sonnet). Finds the project's biggest weakness, challenges planner claims,
|
|
|
|
|
and generates evidence through explore/exploit decisions:
|
|
|
|
|
|
|
|
|
|
- **Explore** (low confidence) — file a `prediction/unreviewed` issue for
|
|
|
|
|
the planner to triage
|
|
|
|
|
- **Exploit** (high confidence) — file a prediction AND dispatch a formula
|
|
|
|
|
via an `action` issue to generate evidence before the planner even runs
|
|
|
|
|
|
|
|
|
|
The predictor's own prediction history (open + closed issues) serves as its
|
|
|
|
|
memory — it reviews what was actioned, dismissed, or deferred to decide where
|
|
|
|
|
to focus next. No hardcoded signal categories; Claude decides where to look
|
|
|
|
|
based on available data: prerequisite tree, evidence directories, VISION.md,
|
|
|
|
|
RESOURCES.md, open issues, agent logs, and external signals (via web search).
|
|
|
|
|
|
|
|
|
|
Files up to 5 actions per run (predictions + dispatches combined). Each
|
|
|
|
|
exploit counts as 2 (prediction + action dispatch). The predictor MUST NOT
|
|
|
|
|
emit feature work — only observations challenging claims, exposing gaps,
|
|
|
|
|
and surfacing risks.
|
2026-03-21 12:44:23 +00:00
|
|
|
|
|
|
|
|
**Trigger**: `predictor-run.sh` runs daily at 06:00 UTC via cron (1h before
|
|
|
|
|
the planner at 07:00). Guarded by PID lock (`/tmp/predictor-run.lock`) and
|
|
|
|
|
memory check (skips if available RAM < 2000 MB).
|
|
|
|
|
|
|
|
|
|
**Key files**:
|
|
|
|
|
- `predictor/predictor-run.sh` — Cron wrapper + orchestrator: lock, memory guard,
|
|
|
|
|
sources disinto project config, builds prompt with formula + Codeberg API
|
|
|
|
|
reference, creates tmux session (sonnet), monitors phase file, handles crash
|
|
|
|
|
recovery via `run_formula_and_monitor`
|
2026-03-23 13:56:59 +00:00
|
|
|
- `formulas/run-predictor.toml` — Execution spec: two steps (preflight,
|
|
|
|
|
find-weakness-and-act) with `needs` dependencies. Claude reviews prediction
|
|
|
|
|
history, explores/exploits weaknesses, and files issues in a single
|
|
|
|
|
interactive session
|
2026-03-21 12:44:23 +00:00
|
|
|
|
|
|
|
|
**Environment variables consumed**:
|
|
|
|
|
- `CODEBERG_TOKEN`, `CODEBERG_REPO`, `CODEBERG_API`, `PROJECT_NAME`, `PROJECT_REPO_ROOT`
|
|
|
|
|
- `PRIMARY_BRANCH`, `CLAUDE_MODEL` (set to sonnet by predictor-run.sh)
|
|
|
|
|
- `MATRIX_TOKEN`, `MATRIX_ROOM_ID`, `MATRIX_HOMESERVER` — Notifications (optional)
|
|
|
|
|
|
|
|
|
|
**Lifecycle**: predictor-run.sh (daily 06:00 cron) → lock + memory guard →
|
2026-03-23 13:56:59 +00:00
|
|
|
load formula + context (AGENTS.md, RESOURCES.md, VISION.md, prerequisite-tree.md)
|
|
|
|
|
→ create tmux session → Claude fetches prediction history (open + closed) →
|
|
|
|
|
reviews track record (actioned/dismissed/watching) → finds weaknesses
|
|
|
|
|
(prerequisite tree gaps, thin evidence, stale watches, external risks) →
|
|
|
|
|
dedup against existing open predictions → explore (file prediction) or exploit
|
|
|
|
|
(file prediction + dispatch formula via action issue) → `PHASE:done`.
|
2026-03-21 12:44:23 +00:00
|
|
|
The planner's Phase 1 later triages these predictions.
|
