scrip-harness 0.8.0

Runnable AGENT.md compile loop for scriptorium: drives `scrip` via subprocess + a model provider. Not part of the deterministic scrip core.

scrip-harness — the runnable compile loop

The deterministic scrip keeper does staleness, provenance, and queries. It never calls a model. scrip-harness is the optional judgment layer that makes the AGENT.md COMPILE step runnable: it asks a configured model provider to synthesize a wiki page from a source, then hands every verifiable step back to scrip.

The dependency points one way only: the harness depends on scrip (and the provider client in model.py); scrip depends on neither. Removing this directory leaves a fully valid, fully deterministic vault and CLI behind.

How a compile runs

scrip-harness compile <slug> (for vault/raw/<slug>.md; pass --from raw/a,raw/b to synthesize one page from several sources):

1. Draft — the selected provider returns a DraftPage: a title, markdown prose with footnote markers [^a1], [^a2], …, and one verbatim quote per marker — each tagged with the source_id it was copied from when several sources are given.
2. Mint + retry — each quote goes through scrip anchor, which rejects a quote that isn't present in the source or isn't unique. Rejected quotes go back to the model for one correction per failure (re-copied or lengthened until unique); bounded retries, then the compile fails cleanly. A hallucinated or paraphrased quote cannot get past this step. Unlike EXTRACT, every claim is kept — the body's [^a1]..[^aN] markers are positional — so a quote is corrected, never dropped.
3. Scaffold + fill — scrip new writes the frontmatter; the harness fills the body with the prose + the minted footnote definitions.
4. Stamp + verify — scrip stamp records provenance hashes; scrip verify proves every citation resolves. If verify fails, the compile errors out rather than leaving a stamped-but-broken page.

So the model owns what to say; scrip owns what is true on disk.

How an extract runs

scrip-harness extract <slug> (for vault/raw/<slug>.md):

1. Draft — the selected provider returns a DraftExtraction: structured claims, each with a verbatim quote, a subject/predicate/object triple, and a polarity.
2. Mint + append — the claims go to scrip fact add --stdin, which verifies every quote (minting anchors), assigns ids and timestamps, skips exact duplicates, and appends all-or-nothing under the write lock.
3. Retry — if quotes come back BROKEN/AMBIGUOUS, the failures go back to the model for one replacement per failure (lengthened until unique, or an empty quote to drop the claim); bounded retries, then the extract fails cleanly.
4. Stamp + verify — scrip stamp vault/facts/_meta.yaml, then scrip verify; contradiction candidates from scrip query contradictions are surfaced for the operator to RECONCILE per AGENT.md.

How a promote runs

scrip-harness promote <slug> (for a compiled vault/wiki/<kind>s/<slug>.md):

1. Score — scrip similar ranks existing pages by overlap (shared sources + title tokens + derived tags) with the candidate, excluding itself.
2. Band the top score: ≥ --merge-threshold (0.5) → merge into it, deterministically (no model); < --keep-threshold (0.25) → keep the page as its own; in between → the model decides merge-vs-keep over the small candidate set (the only model call in PROMOTE).
3. Merge — append the candidate into the target (its [^a1].. footnotes renumbered to avoid collision), union the derived-from, record the absorbed id in supersedes, delete the absorbed page, then scrip stamp + scrip verify. --dry-run prints the decision and mutates nothing.

How an answer runs

scrip-harness answer "question" makes the ANSWER rung executable:

1. Preflight — scrip status, scrip verify, and scrip query contradictions must be clean by default. Stale artifacts, broken anchors, or open contradiction pairs stop the answer before any model call.
2. Gather — the harness ranks claims from facts/, reads relevant compiled wiki pages as context, and falls back to scrip search when compiled evidence is thin.
3. Draft — the selected provider answers from that bounded evidence packet and returns structured citation records: either an existing claim_id or a verbatim raw quote.
4. Verify citations — claim citations are resolved with scrip span; raw quotes are minted with scrip anchor. Unsupported citations fail the answer. --save writes a verified note under wiki/explorations/.

How a reconcile runs

scrip-harness reconcile (over every open contradiction):

1. Find — scrip query contradictions lists the candidate pairs (same subject+predicate, opposing polarity, different sources, not yet adjudicated).
2. Read — for each pair, scrip span --claim <id> fetches both verbatim cited spans, and the model decides supersede (with a winner), qualify (with a verbatim qualifier quote + the condition under which it holds), or keep-both, with a rationale.
3. Record — the decisions are written append-only with scrip fact add --table reconciliations (existing claim rows are never rewritten); a qualify also authors a polarity: qualifies claim via scrip fact add --table claims (its anchor minted + verified). Logged to wiki/log.md, then scrip stamp + scrip verify. Adjudicated pairs stop being surfaced by scrip query contradictions. --dry-run prints the decisions without recording.

Install & run

Both packages are on PyPI. scrip-harness bundles scriptoria as a dependency and drives it through its own interpreter, so it is self-sufficient — install scriptoria as a tool too only if you want the scrip command on PATH for direct use:

uv tool install scrip-harness            # this package → `scrip-harness` (pulls scriptoria)
uv tool install 'scriptoria[ingest]'     # optional: `scrip` on PATH + HTML/PDF ingest
export OPENAI_API_KEY=...                 # or ANTHROPIC_API_KEY / GEMINI_API_KEY

scrip-harness compile article --provider openai
scrip-harness extract article            # pull claims into facts/
scrip-harness answer "What does the corpus say about caching?" --provider openai
scrip ingest <url> --slug article        # bring a source in (needs the install above)

(From a checkout, uv tool install ./scrip and uv tool install ./harness install the local versions instead.)

Provider selection

Every model-backed command accepts:

scrip-harness answer "..." --provider auto|anthropic|openai|gemini \
  [--model MODEL] [--api-key-file PATH]

--provider auto is the default. It picks the first available key in this order: ANTHROPIC_API_KEY, OPENAI_API_KEY (or ~/veed/var/openai), then GEMINI_API_KEY/GOOGLE_API_KEY (or files under ~/veed/var/gemini). Provider defaults can be overridden with SCRIP_HARNESS_<PROVIDER>_MODEL. Environment keys take precedence over key files. When --api-key-file points at a directory, the harness reads the first non-empty key from the sorted files in that directory. An explicit key file also needs an explicit provider:

scrip-harness answer "How does the answer ladder work?" --provider openai \
  --api-key-file ~/veed/var/openai
scrip-harness answer "How does raw fallback work?" --provider gemini \
  --api-key-file ~/veed/var/gemini --model gemini-3.5-flash

Demo fixture

From a checkout, scripts/demo_answer.sh runs the full answer command against a small synthetic vault in examples/answer-demo-vault/:

scripts/demo_answer.sh --provider openai
scripts/demo_answer.sh --provider gemini --api-key-file ~/veed/var/gemini
scripts/demo_answer.sh --root . --provider auto "What does the vault say about caching?"

The script runs scrip status and scrip verify before invoking the model. Pass --save to write the verified answer under vault/wiki/explorations/.

Develop / test

cd harness && uv run pytest        # hermetic: the model is stubbed; scrip runs for real

The tests inject a stub draft function (no network, no API key) and drive the real scrip subcommands over a temp vault, asserting the result is stamped and verified.

Scope & limits (v1)

• Covers COMPILE (one or more sources → one wiki page, with the bounded quote-retry loop), EXTRACT (one source → claims in facts/, same retry loop), ANSWER (fresh compiled evidence first, raw search on miss, verified citations), PROMOTE (score → merge/keep, model only in the middle band), and RECONCILE (adjudicate every contradiction → record the decision). Entities/edges go through scrip fact add --table entities|edges by hand.
• COMPILE accepts one or more sources (--from raw/a,raw/b); EXTRACT is still one source. PROMOTE's merge is append, not re-synthesis — multi-source COMPILE now unblocks re-synthesis as a follow-on. reconcile records the decision (supersede/qualify/keep-both) and, on a qualify, authors the nuancing polarity: qualifies claim; surfacing the page caveat is left to the read-only view layer, not a page mutation.