Mirofish codex

MiroFish Codex-native

MiroFish is a local-state, schema-driven runtime beneath the current Codex or Claude session. The current coding agent is the only LLM harness.

Hard boundary

Never start codex exec, claude -p, an OpenAI-compatible generation endpoint, or an Ollama generative model from inside the production workflow. Doing so nests harnesses and changes the model being evaluated.

The project CLI may load a local multilingual embedding model. Embeddings are used only for retrieval and similarity scoring; they never generate agent actions, reports, or prose.

Choose a workflow

Intent	Command path
Inspect setup	mirofish-native doctor --quality
Ingest and simulate	init → step / resume loop
Search local memory	search, neighbors, or timeline
Build a report	continue the same run through report reflection
Novel N+1 benchmark	prepare → anonymize submissions → freeze → reveal target → score → export-public → reveal mapping

Run commands from the repository's backend directory with uv run --no-dev mirofish-native ... unless the console script is already installed.

Start a native run

uv run --no-dev mirofish-native init \
  --workspace /absolute/private/run-directory \
  --source /absolute/source-file \
  --question "要推演的问题" \
  --local-files-only \
  --locale zh

The workspace contains private source-derived artifacts and must be outside every Git worktree; an ignore rule is not sufficient. --local-files-only persists an offline embedding policy for every later command; omit it only when the pinned model may be downloaded on first indexed use.

Execute the reasoning loop

1. Call step --workspace ... --json.
2. If the status is completed or failed, stop.
3. If the status is needs_reasoning, treat the inline task object returned by step as the immutable task packet and retain its context_sha256.
4. Produce exactly one response envelope matching its JSON Schema. Do the reasoning in the current top-level session; do not delegate to a nested CLI.
5. Save the response as UTF-8 JSON and call:

uv run --no-dev mirofish-native resume \
  --workspace /absolute/private/run-directory \
  --response /absolute/private/response.json \
  --json

6. On schema failure, repair the same task. Do not invent missing fields, skip the task, or mark partial JSON successful.
7. Continue until the CLI reports a terminal state.

If initialization was interrupted before active-run.json was updated, call list-runs --workspace ... --json and resume with an explicit --run-id.

For each run, the broker leases at most one active task at a time. Top-level subagents, if the host supports them, may collaborate only within that active task. They must not call step or resume independently, acquire another packet, or submit competing envelopes. The top-level session returns exactly one response envelope before the broker can lease the next task.

Read references/protocol.md before changing task packets or implementing a new workflow phase.

Use local evidence tools

When a task allows evidence tools, call only the named local command and use the returned run-scoped JSON while solving that task:

uv run --no-dev mirofish-native search --workspace ... --query "..." --limit 12 --json
uv run --no-dev mirofish-native neighbors --workspace ... --entity-id "..." --depth 2 --json
uv run --no-dev mirofish-native timeline --workspace ... --entity-id "..." --json

Do not substitute a web search for private run memory. Do not allow one run's vectors or graph nodes to enter another run's candidate set. Evidence queries do not change the dispatched task identity; echo its original context_sha256.

Quality profile

For quality-sensitive work:

• require BAAI/bge-m3@5617a9f61b028005a4858fdac845db406aefb181;
• use dense + Chinese character-aware lexical + graph retrieval;
• keep entity aliases, ambiguity, provenance, fact validity, and supersession;
• fail closed if dense indexing is incomplete;
• finish the memory-write drain before report reads;
• run explicit report draft → critique → revise rounds;
• retain exact context hashes and model identity where exposed; otherwise record the literal not_exposed.

Never silently fall back to a 7B/32B generation model. If the current Codex/Claude session cannot continue, leave the run resumable and report the blocker.

Novel benchmark

Treat the source and generated chapters as private, user-supplied material. Do not commit source prose, target prose, or full model continuations.

Read references/novel-benchmark.md, then:

1. pre-register the parser version, N, candidate runtime settings, fixed rubric, and randomisation-seed commitment, then prepare separate context/target artifacts;
2. freeze every candidate before revealing chapter N+1;
3. anonymise labels with a separately stored mapping;
4. extract roughly ten target plot nodes after freeze;
5. judge strict node hits and anonymous 0–100 similarity;
6. compute BGE-M3 cosine as an independent anchor;
7. use benchmark export-public for aggregate scores and hashes, then review any separately transcribed runtime configuration;
8. reveal model labels after scores are immutable.

MiroFish outputs are non-canon candidates. They do not update an author's story state unless the author explicitly accepts them.