Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.
Backend capability and limit authority audit
Lane B, Session seshneu2u9cbw, base
76e269c9a9dfd3a6286ce0da81b77316ca4be38a.
Contract
Unknown custom capability metadata must not suppress supplied images or
reasoning. Explicit modality lists, supports_reasoning: false, and effort
lists retain their meanings. Native rows keep their own metadata. Custom rows
must not acquire an unrelated model's context, tools, service tier, or default
effort. Persisted BackendModel shapes and Avibe-owned model selection remain
unchanged.
Authority and ambiguity inventory at the shared base
- BackendModel stores optional context/output numbers, optional capability
booleans, and empty-by-default modality/effort lists.
originidentifies the creation path for the whole row; it cannot distinguish an imported field from a later user edit. No existing per-field limit provenance exists. - The UI's candidate path leaves limits and capabilities unstated. Its manual editor visibly starts with text input and reasoning disabled. A models.dev selection fills the editor; saving commits that full row. The backend saves those values literally and built-in reconciliation only adds missing rows. Therefore a saved non-null limit should remain the selected model's description, without trying to reconstruct how it was authored.
- Route resolution reads metadata by the requested BackendModel ID, before mapping it to the upstream target. This is intentional: an alias owns its planning metadata. Codex's authenticated turn gateway retains that alias as the runtime model, so the projected catalog row matches. Claude launches the target but carries the selected row's limits. OpenCode overlays use the public alias and selected row's limits.
- Claude overwrites both
CLAUDE_CODE_MAX_*_TOKENSenvironment variables, in both Hub and native CLI launches. Hub also masks them as if they were credentials. Separately,claude_settings_for_launchcopies the catalog limits into the highest-priority launch settings, so changing only the subprocess environment would still lose explicit choices. Project/local settings are enabled, user settings are disabled for Hub transport ownership. Native CLI retains all setting sources. Parent environment, settings env, and the saved model row can conflict; there is no provenance for ranking two user-authored descriptions. Credential/connection masking is independent. - Codex projects the saved context into
context_windowandmax_context_window, removing the old catalog compaction limit so it can be derived from the new context. Custom rows without a limit inherit no context. Launch arguments do not inject a context override. Native CLI launches do not consume this projected catalog. The real consumer checks below covermodel_context_windowandmodel_auto_compact_token_limit. BackendModel output limits have no Codex projection; silently claiming support or inventing a new output cap would be incorrect. - OpenCode projects saved limits into
limit.context/limit.outputwhen present and otherwise omits them. The managed plugin replaces complete Hub provider definitions to protect authentication, so a same-name native provider cannot remain a second owner. Ordinary native provider definitions are separate. No new merge of arbitrary native provider data is proposed. - Unknown input metadata is converted into
["text"]only for custom Codex rows. Codex 0.153.2's protocol schema itself defaults omitted input modalities to text and image. Explicit nonempty lists can narrow that default without adding persisted state. Original-resolution image detail is a separate provider feature and should not be enabled from absence. - Unknown reasoning is converted into a
none-only list and default. Explicit false and explicit lists must remain distinct from absence. - The UI model picker currently returns an empty effort menu for an unknown row, and the manual editor starts with explicit reasoning false. These are separate visible-product policies; this lane will not change them without coordination.
Approved repair
The orchestrator approved the limit rule in integration commit dc2d5f7bc and
the Codex consumer follow-up in a3065a7d1. The shared plan remains owned by
the orchestrator.
- Preserve the persisted schema and requested alias's metadata authority.
- Unknown custom Codex input omits
input_modalities, invoking the native text/image compatibility default, while original-resolution detail remains disabled. A nonempty modality list still narrows the accepted inputs. - Unknown custom reasoning has the schema-required empty supported list and
no default effort. Actual turns can still supply an effort. Explicit false
retains
none, explicit lists retain their selected/default efforts, and known native rows retain their own defaults. - Claude catalog limits fill only absent subprocess variables, including preservation of explicitly empty strings. Only the two limit keys leave the credential tombstone set. Hub connection settings no longer promote catalog limits above native project/local settings. Credential masking and enabled settings sources are unchanged.
- Applying a saved Codex context now removes
max_context_windowinstead of synthesizing an equal hard ceiling. The planning context remains and the obsolete native compaction threshold is removed as before. Without a saved context override, native model metadata remains unchanged. - OpenCode needs no production change: the saved BackendModel remains the managed provider's authority. Alias/overlay tests confirm supplied limits and omitted metadata do not inherit an unrelated target row's values.
Consumer evidence and limitations
- Codex 0.153.2 loaded a native context of 512,000 but originally reported only
121,600 usable tokens with the projected 128,000 hard ceiling. The pinned
source's
models-manager/src/model_info.rs::with_config_overridesconfirms the minimum operation. Removing the synthesized ceiling makes its actual token-usage consumer report 486,400 usable tokens, preserving 95% headroom. Smaller explicit windows and the catalog fallback also work. - Two-turn tests feed real CLI responses with synthetic token usage. With the 128,000 planning window, 150,000 tokens trigger compaction. With an explicit 512,000 window and 300,000 threshold, 150,000 do not trigger compaction and 350,000 do. The mock sees the extra compaction request and the app-server completes its compaction item and following turn.
- Actual Codex requests retain a valid inline PNG and non-ASCII text. Unstated
effort adds no default; explicit
none,minimal,low,medium,high,xhigh, andmaxreach the local endpoint. Negative image/reasoning choices and nonempty effort lists are exercised through the same consumer. - Codex
ultrais a native multi-agent mode: the pinned client'sreasoning_effort_for_requestmaps it to a non-Ultra tier, falling back to medium for an empty ladder. This repair does not change that native interpretation or claim literal Ultra wire preservation. Native CLI tools may also have their own defaults independent of catalog projection. - Claude SDK 0.2.93's bundled CLI preserves parent/project/local output limits
in real Messages requests, for both Hub and native CLI paths. Its normal
context planner does not consume
CLAUDE_CODE_MAX_CONTEXT_TOKENSunless compaction is disabled. The repair preserves that value through SessionHandler into SDK options but does not claim to change native Claude context planning. Compaction is never disabled as a workaround. - BackendModel output metadata has no Codex output-cap projection. No output cap, UI policy, schema, engine, gateway, credential, or routing-policy change is introduced. Engine-to-upstream preservation remains lane D's boundary.
All CLI tests use test-owned homes, allowlisted environments, synthetic tokens, and loopback endpoints. External HTTP(S) is directed at a test-owned rejecting proxy. On macOS, a separate OS sandbox denies non-loopback outbound sockets; a representative forbidden connection was verified to return a permission error. Claude nonessential traffic is explicitly disabled; Codex launches use Avibe's existing analytics/update/host-feature disable settings.
Validation results
- Baseline: 70 catalog/injection tests passed before the repair.
- Focused catalog, injection, and SessionHandler coverage: 201 passed.
- Related scenario/catalog/route/projection selection: 219 passed.
- Actual Codex/Claude consumers: 30 cases, including the three compaction boundaries above.
- Pinned Ruff 0.4.9 and Model Hub authority closure pass.
- These are implementation-head results, not evidence for a later PR head. The owner's 2026-09-09 independent-delivery decision assigns this lane its own PR, canonical CLI fixture registration, and durable review/CI Watch. MH-PROTOCOL-004 remains partial CLI-only evidence; MH-CLAUDE-LAUNCH-001 gains related launch-limit evidence without claiming managed-engine coverage. After normal landed-master integration, rerun focused tests, real isolated CLI consumers, authority/scenario validation, and pinned Ruff at the PR head. The orchestrator independently verifies all gates and executes any merge; no runtime update or publication is authorized.