deepreason-orchestrator git:20260805.367ad6aA

Immutable. This exact content is served forever at /api/v1/blob/591e8bb705cf5cf4.

---
name: deepreason-orchestrator
description: Entry point for any DeepReason problem. Routes work to exactly one subskill at a time and enforces the scope contract. Use when asked to manage, diagnose, or fix anything in DeepReason.
---

# DeepReason problem orchestrator

You are running a tightly bounded workflow. You do not freelance. You
select ONE subskill, execute it to its exit criteria, then return here
to select the next. You never blend phases.

## The scope contract (read before every phase)

1. **One tranche = one goal.** A tranche works exactly one GOAL.md
   (produced by `dr-set-goal`). Anything else you notice goes into
   `PARKED.md` — never into your work. Write the parked entry for its
   FUTURE RUNNER, at park time, while the context is free: one line of
   WHAT, then a ready-to-send prompt (route, one-goal statement,
   evidence pointers, end state). Starting the follow-up should cost
   the operator a paste, not an authoring session.
2. **Evidence over prose.** Claims about DeepReason behavior are only
   admissible if derived from typed records: `log.jsonl`, `objects/`,
   `progress.jsonl`, `run-status.json`, `REPLAY_VALIDATION.json`,
   `verify_root`, or test output. Your own summary of what code
   "probably does" is not evidence.
3. **No phase-skipping.** You may not implement without an approved
   FIX.md. You may not write FIX.md without a DIAGNOSIS.md. You may not
   write DIAGNOSIS.md without a reproduction or record-derived trace.
4. **Stop conditions.** Stop and report (do not improvise) when: a
   command fails twice the same way; evidence contradicts the goal;
   the fix requires touching frozen-record semantics (anything under
   `capabilities/state.py` digests, `harness.py` event application, or
   replay validation record formats); or the diff would exceed ~150
   changed lines. Before any stop becomes a question to the operator,
   load `dr-ask-the-right-question`: route it to the cheapest authority
   first, and ask only what survives the dominance test — batched, with
   a recommendation.

## Map preflight (do this before routing, every time)

`docs/map/` is the navigation layer over 125k lines of source. Scoping
from grep instead of from the map is how a change misses a call site.

1. Read `docs/map/INDEX.md` and resolve the work to ids:
   `DR-SUB-<pkg>`, `DR-CON-<concept>`, `DR-SEAM-<a>-x-<b>`.
2. If the work spans two things, **read the SEAM document first**. It
   says which fraction of each side is actually involved, which is
   usually small. Reading both subsystem documents first is reading ten
   times more than you need. The file is `docs/map/SEAM-<a>-x-<b>.md`,
   sides in alphabetical order; the worked recipe for changing one is
   `docs/map/REC-change-a-seam.md`.
3. Read `docs/map/INV-frozen-surfaces.md` BEFORE designing anything.
   Discovering a frozen surface after the code is written is the
   expensive order to discover it in.
4. Record the resolved ids in the tranche's first artifact (GOAL.md or
   REQUEST.md). Every later phase starts from the same map.

If the map has no id for something the work touches, that is a finding,
not a blocker: say so, and creating the missing document becomes part of
the tranche. `docs/map/SCHEMA.md` is the contract for writing one.

The map is maintained by the phases that change code, in the same
commit — see `dr-execute-step` and `dr-implement-fix`. Nothing else may
advance a `Verified-at:` stamp.

## Environment preflight (run once per session, before routing)

The full driving manual — preflight, CLI lifecycle, ladders, where to
look — is `dr-drive-harness`; load it if this session has not run the
harness before. The cloud container rolls back silently. Verify, in order:

    git log --oneline -1                # expected branch head, not stale
    git status --porcelain | head       # know what is uncommitted
    which deepreason || pip install -e . --break-system-packages -q
    ls experiments/live_research_*/env  # credential file survives rollback? if listed in the goal, recreate per its README/handover

If anything was stale: resync the working branch
(`git fetch origin <branch> && git checkout -B <branch> origin/<branch>`)
before any other action.

## Routing table

Ask: what does the current state of the tranche lack?

| Missing artifact | Route to |
|---|---|
| No GOAL.md for this tranche | `dr-set-goal` |
| GOAL.md exists, no DIAGNOSIS.md | `dr-diagnose` |
| DIAGNOSIS.md names a cause but nothing demonstrates it | `dr-reproduce` |
| Reproduction exists, no FIX.md | `dr-propose-fix` |
| FIX.md exists, code unchanged | `dr-implement-fix` |
| Code changed, outcome unverified | `dr-verify-outcome` |
| dr-verify-outcome reported PASS | Tranche complete: report and stop |

If `dr-verify-outcome` reports FAIL, route back to `dr-diagnose` with
the failure evidence appended — do not patch forward from intuition.

## Tranche working directory

All tranche artifacts live in one directory the goal names, e.g.
`experiments/<date>-<slug>/`: `GOAL.md`, `DIAGNOSIS.md`, `REPRO.md`,
`FIX.md`, `VERIFY.md`, `PARKED.md`. Commit and push this directory at
every phase boundary (the container can vanish at any time):

    git add -A <tranche-dir> && git commit -m "<phase>: <one line>" \
      && git push -u origin <branch>

## Hard prohibitions (apply to every subskill)

- Never modify a committed run root's contents. Run roots are retired,
  never edited: `git mv run-<id> <state>-epochN-run-<id>` and commit
  the rename BEFORE any relaunch (deterministic identity otherwise
  refuses with RUN_ALREADY_STARTED).
- Never commit credential material. `env` files are gitignored; check
  with `git check-ignore` before writing near them.
- Never run the full live ladder to test a code hypothesis — that is
  `dr-verify-outcome`'s final step only, and only when the goal calls
  for live proof.
- Never widen the goal because the codebase "needs" it. PARKED.md.