# Model-led runtime

Omnius uses a short, stable core prompt and lets the main model choose its plan, tools, and response within the current user's scope. A request for a proposal remains a proposal until the user authorizes implementation.

The core prompt is 977 characters of project-specific collaboration guidance. The [official Codex prompting guide](https://developers.openai.com/cookbook/examples/gpt-5/codex_prompting_guide) was a design reference. The user's [third-party prompt example](https://github.com/asgeirtj/system_prompts_leaks/blob/main/OpenAI/Codex/gpt-6.1-sol.md) has unverified authenticity. Its claimed provenance is not established, and no safety, app, or plugin boilerplate was imported into the Omnius prompt.

## Request flow

The runner assembles the current request from the core prompt, explicit project/user instructions, the conversation, authenticated tool results, and current task facts. The late `RUN OBSERVATIONS` block contains JSON records such as attempted calls, recent results, changed paths, declared todos, and verification receipts. It does not select the next operation.

Automatic workflow scaffolds, focus directives, edit pressure, repetition interventions, unsolicited critiques, and automatic recovery advice are retired by a closed runtime policy. Legacy configuration cannot reactivate them. Existing checkpoints retain their source records; recognized old host-control wrappers are retired from model context rather than rewritten into a new instruction.

Observation renderers distinguish a host-supplied semantic outcome digest from a digest of display text. Archive handles or presentation changes alone therefore do not justify a claim that the source changed. Recorded execution status, source revisions, read ranges, fidelity, and explicitly omitted excerpts remain visible. Text within a source is preserved as data, including text that resembles an instruction.

Inferred `current_plan` and `constraint` memories remain queryable historical records. Their classification, chat key, importance, or TTL cannot pin them as current system instructions. Retrieval follows immutable source provenance and scope; original records remain available. The current user's request and actual task lineage supply current authority.

The capability registry describes available names, schemas, and exposure. A missing-tool receipt may name related registered tools, but it does not activate or invoke an alternative. A model-selected, schema-valid direct tool request and explicit tool discovery remain supported.

Telegram's admitted execution context describes capabilities and delivery fields without prescribing browser/media tools, delegation, mandatory progress calls, or a writing standard. Skills remain available on demand. Todo records are factual JSON. No-match counts, objective plateaus, or a ready-to-deliver classification cannot forbid the model's next operation.

Cache-preserving updates carry complete changed observation snapshots. They do not tell the model to interpret deleted lines as facts that became false. Optional background summaries defer while the main recurrent-model prompt cache is reserved; required task work retains an explicit cache handoff.

Automatic embedding-based selection of failure-recovery advice is retired with that advice. Historical pattern records remain intact, and explicit memory tools remain available. Ordinary startup does not load an embedding model to rank retired hints.

On cold resume, completed historical tool transactions become quoted observation records rather than replayed assistant behavior. File-read results and other source records remain inline. Completed shell, grep, directory, and file-discovery diagnostics may instead carry hashes and immutable archive references; their exact arguments and output, including any useful source text, remain readable through the real `artifact_read` tool. A failed archive write retains the full inline record. Visible assistant notes stay intact, and the exact original transcript is durably archived. Only exact duplicates are grouped. Ambiguous, unanswered, protected-source, and new transactions retain their native protocol representation.

Known old host presentations—recalled notes, procedures, passive chat, and strictly identified historical receipt frames—are archived with their complete original metadata and byte hashes. Mixed frames containing goals, constraints, project instructions, or unknown sections stay intact. Telegram's current host context is rebuilt separately from its archived prior presentation. These are data representations, not instructions to choose another tool or finish a task.

A model-authored approval question under an inferred delivery contract can retain its exact plan reply without inventing a durable input-suspension record. Authored input criteria still require their exact IDs and a committed suspension. The Telegram reply carrier continues to require the current run, epoch, objective scope, and terminal receipt.

## Boundaries retained

Authorization, caller-owned task requirements, source/hash integrity, effect receipts, provider protocol, cancellation/Stop, physical capacity limits, and scoped storage remain enforced. They establish which effects are allowed and what actually occurred; they do not invent a behavioral plan.

## Validation and limits

CPU regressions exercise recorded source retention, semantic/display identity, factual rendering, historical memory scope, actual outbound prompt assembly, and capability recovery with a fake transport. They cannot establish the quality of an unobserved model response or prove production delivery.

The final focused run passed 847 orchestrator cases. Telegram/web lifecycle checks passed 174 cases; the final restart/native-delivery subset passed 68, overlapping those lifecycle checks. The live harness passed 19 CPU cases, including byte-equivalent metadata-derived identifier sanitization.

Real inference used `robit/qwen3.8-27b-obliterated-e03:27b` on a verified A100, with context 262144, output limit 150925, temperature 0.3, and thinking disabled. All three cases used the same frozen production runtime and a predeclared 16-request limit:

| Case | Primary requests | Real inspections/work tools | Observed result |
| --- | ---: | ---: | --- |
| Captured-history planning | 15 | 23 | Accepted grounded plan; no project edits; approval boundary retained |
| Fresh planning | 10 | 33 | Accepted grounded plan; no project edits; approval boundary retained |
| Controlled implementation | 7 | 8 | Real source/test edits; model ran `npm test`, 4/4 passed; independent behavior and test verification passed |

The fresh planning trace made 22 real `file_read` calls. Every earlier source body remained byte-for-byte present through all subsequent requests, including the final request. The captured-history case retained all 42 original observations: 41 diagnostic calls had exact archive readback and one progress result remained inline. Its original capture contained no `file_read` calls; inline file-read retention is additionally exercised by the genuine cold-runner test with actual ShellTool, FileReadTool, and built-in artifact retrieval.

Request-prefix extensions and full rebases both occurred. Required frame-size rebases remain bounded, and the live trace had no auxiliary model calls; these observations do not establish a guaranteed cache-hit rate. Hashes, per-case checks, and limitations are recorded in [model-led-validation.json](model-led-validation.json). Private goals, source fixtures, and chat dumps remain local.

A full-suite pass has not been established. The broad updated-runtime run was interrupted with 93 failing test files. Legacy tests that require retired behavioral controllers need different expectations; failures at authority, source-integrity, and delivery boundaries require comparison against the same baseline before being classified. Follow-up is tracked in [issue #17](https://github.com/robit-man/open-agents/issues/17).

For this change, five selected authority/source/delivery suites produced identical baseline and updated outcomes: 96 cases, 68 passing and 28 existing failures, with no changed test outcomes. This comparison does not establish full-suite health.

The broad Telegram Bot API suite was also compared using the same private project fixtures on both versions: 132 passing and 32 failing on each, with identical failing test names. Its former checkout-based fixture loaded the operator's project credential references into an empty test vault; private fixture roots remove that contamination without changing credential validation.

The early captured-state replay against the prior runtime **passed** with a narrower tool set than the final harness. It did not reproduce the reported baseline stagnation, so the change is not supported by a claim of a reproduced failing baseline. Later replay with directives retired but the full old diagnostic history still inline failed after 16 requests, including eleven consecutive identical conference greps. The revised representation archives those completed diagnostics and old host presentations while retaining exact retrieval custody. Live outcomes must be reported with their model/runtime, request artifacts, observed behavior, and remaining limits. A passing replay demonstrates that run's behavior; it does not guarantee that future model choices will avoid every loop.
