Runtime Algorithms and Direction

September 17, 2026 · View on GitHub

Current feature status (audit §19.3)

"Implemented", "wired", "enabled", "verified in CI", "released", and "measured benefit" are different gates. A pure function passing unit tests is not a live product path, and a live path is not a measured improvement. This table records each mechanism's actual gate at the pinned commit; the dated baseline below is history, not current truth.

Mechanism	Implemented	Wired into live path	Enabled by default	Verified in CI	Released	Measured benefit
DAG claim scheduler (`tool-dag-scheduler`)	yes	`agent-loop` frontier executor	`toolScheduler: "dag-v2"`	unit + integration tests	v0.99.x line	not measured
Ready-frontier admission (`runDagFrontier`)	yes	`agent-loop`	same	`tool-dag-ready-frontier`, `tool-dag-hook-replan` tests	working tree	not measured
Dynamic-claim conflict check (`conflictsWithUnsettledClaim`)	yes	`agent-loop`	same	`tool-dag-dependencies` + hook tests	working tree	not measured
ECRAF admission planner (`tool-dag-ecraf`)	yes	no — no call path wires it	n/a	unit tests only	unreleased	not measured
Context Budget V2	yes	system-prompt assembly	opt-in policy	planner/selection/cache tests	working tree	not measured
Tier floor reservation	yes	context-budget-v2 planner	when V2 enabled	`context-budget-v2-tier-floor` tests	working tree	not measured
Reasoning router v4	yes	`/think auto` lane	opt-in	router tests	released	classification only, not success-probability calibration
Workload permit pool	yes	resource admission	default	pool tests	released	not measured
Provider retry/failover classification	yes	`provider-retry`, `session-failure-cause`	default	resilience/classification tests	released	not measured
MCP descriptor injection screen	yes	`mcp/manager` import path	default	quarantine tests	released	pattern rule score, not calibrated risk
verified-run coordinator + evidence	yes	verified-run paths	opt-in command	coordinator/evidence tests	working tree	scope-limited binding, not general correctness

Gates are reported per row so a green "implemented" never upgrades itself to "released" or "measured".

ECRAF arithmetic boundary (2026-09-17)

The internal planEcrafAdmissions() planner rejects non-finite derived density denominators, scores, and reserved usage with RangeError, even when each input number is finite. Scores are computed once per candidate before sorting or calling the conflict predicate; singleton and zero-slot batches receive the same validation. A bounded resource overflow still defers the candidate and allows later feasible candidates. An unbounded resource has no capacity limit, but its usage must remain representable as a finite number. Failed passes return no partial admission plan and do not mutate caller input.

Regression coverage is in packages/agent/test/tool-dag-ecraf-arithmetic.test.ts; existing numeric and property tests remain in packages/agent/test/tool-dag-ecraf.test.ts. This is local pure-planner validation, not live frontier wiring, CI confirmation, a release, or evidence of latency/quality improvement. Resource normalization, slot-aware ranking, fairness, and equal-budget runtime comparisons remain separate work.

Working-tree shared run budgets

The SDK prompt(..., { runBudget }) path now shares a monotonic deadline and logical request/concurrency limits across the active prompt's main stream, retries, continuations, and first-party summaries using that stream. Exhaustion is a non-retryable budget_exhausted termination; snapshots keep outstanding streams until terminal metadata arrives. No request/time budget is imposed by default. Preflight is now owned even for unbounded prompts; unresolved streams block a later prompt instead of being discarded when their budget scope closes. See Shared run budgets for units, zero semantics, cancellation, and uncovered paths. This is not billing enforcement, a persisted budget, or a hard process-termination deadline.

Working-tree execution ownership (2026-09-10)

The live session now retains registered tool-promise ownership after timeout or abort, defers prompt settlement/resource-lease release until actual termination, and rejects overlapping prompt admission. Shared permits capture request values, honor zero capacity, and wake FIFO followers after head cancellation/expiry. The internal lane launcher forwards cancellation and respects zero/heavy caps. Independent bash calls now own separate cancellation controllers, and one call's completion cannot hide another active call. Core teardown waits use a monotonic clock, preserving the existing grace interval across wall-clock changes. See Prompt settlement for tests, compatibility, and uncovered paths. These changes do not implement a durable verified-run product.

v0.98.3 release delta (2026-09-06)

The explicit advisory SDK requires normal first-party model completion, honors cancellation, reports caller-rank top ties, and preserves intake/missingness diagnostics. See Advisory selection integrity. No default AgentSession/TUI judge, additional completion calls/retries, or calibrated risk policy is activated.

omk-protocol also exports Claim Closure Graph v1 evaluation; omk-adaptorch-wpl exports proof-result/VERA projections. They consume caller-supplied evidence, not authenticated runner truth. The operation-trace and Effect Journal V2 modules in omk-agent-core remain internal primitives: this release does not wire them into a new live authority loop.

The dated audit below preserves the v0.97.0 baseline and its then-working-tree classifications. The context floors, explicit-rule compactor, resource report, settlement notifications and OpenWiki guards described as working-tree changes below subsequently shipped in v0.98.0. That does not ship a generated OpenWiki corpus or promote the blocked lane/memory policies. Source and tests remain authoritative.

Historical v0.97.0 baseline audit

Snapshot date: 2026-08-27
Released baseline: OMK v0.97.0 (b38a2c8c84)
Repository baseline: 4b79c65eaf plus the local working tree

Status vocabulary

Status	Meaning
Released / default	Present in `v0.97.0` and selected when relevant settings do not override it
Released / opt-in	Present in `v0.97.0`, but requires a command, setting, or explicit API call
Released / internal	Implemented and tested, but not connected to a live user path
Working tree	Present in the current checkout only; not shipped and not a release promise
Proposed	Design direction without a complete implementation and verification path

A mechanism's existence does not make it authoritative. OMK promotes a mechanism only after its live call path, default, evidence, and rollback are all explicit.

Runtime control path

Prompt or durable-goal round
  -> system-prompt assembly
     -> Context Budget V2 when globally enabled
  -> local v4 reasoning routing when /think auto is active
  -> provider attempt
     -> same-family route rotation, retry, or configured failover when eligible
  -> tool scheduling
     -> v0.97.0 CLI default: dag-v2; direct core fallback is revision-specific
     -> pre-hook claims -> deferred authorization -> post-hook claim re-plan
     -> deterministic level execution and source-order results
  -> journals, receipts, and session termination records
  -> optional omk-protocol evaluation
     TaskSpec -> ExecutionAttempt -> Observation -> EvaluationResult -> RuntimeDecision
  -> retry, continuation, or final prompt settlement

The protocol line is a released library and adapter surface. An ordinary chat turn does not automatically become a TaskSpec; callers opt into that semantic evaluation contract.

Current algorithm surface

Tool scheduling and settlement

Status: Released / default. The v0.97.0 CLI sets dag-v2; direct omk-agent-core calls fall back to waves-v1. The working tree changes that core fallback to dag-v2, but that promotion is unreleased. The scheduler canonicalizes resource claims, preserves source order, and places conflicting calls in later levels. Before a level executes, OMK authorizes its calls and re-plans with post-hook arguments so a hook cannot silently invalidate the original claim plan.

The live executor uses level barriers. assignDagDependencies() computes a finer predecessor graph, but no live executor consumes it. Each candidate level is authorized and then re-planned from post-hook arguments.

Evidence:

packages/agent/src/tool-dag-scheduler.ts: assignDagLevels, assignDagDependencies, scheduleDagLevels
packages/agent/src/agent-loop.ts: executeToolCallsDagLevels, runDagLevelCalls
packages/agent/test/tool-dag-scheduler*.test.ts
packages/agent/test/tool-dag-dependencies.test.ts

Working tree: timed-out tools receive a bounded 250 ms teardown window. A cooperative process may settle and let the model react to the timeout; a tool still running after the window stops the run because it may still mutate the workspace. This candidate lives in packages/agent/src/tool-timeout-settlement.ts and is covered by packages/agent/test/tool-timeout-loop-continuation.test.ts.

Context selection

Status: Released / opt-in. Context Budget V2 is enabled globally through contextBudget.enabled or per process with OMK_CONTEXT_GOVERNOR=1.

The planner:

reserves response and safety tokens;
pins hard or required items first;
computes tier floors and ceilings;
scores optional items for relevance, recency, evidence, redundancy, priority, and full-text token cost;
sorts optional items by density -> effectiveScore -> priorityRank -> fullTokens -> id; and
selects a full, summary, headroom-compressed, pointer, or omitted representation that fits.

Density divides effective score by the cheapest non-omit representation (admissibleTokens), not by full-text size. This avoids penalizing an item that can be represented by a small evidence pointer. Stable item IDs and explicit selection policy sel-2 make tie-breaking and plan-cache invalidation deterministic.

When enabled, representation and negative-result entries persist under .omk/cache/context-budget-v2; plan entries remain session-memory-only. OMK_CONTEXT_GOVERNOR_CACHE=memory keeps every cache entry in session memory, and OMK_CONTEXT_GOVERNOR_CACHE_DIR relocates the representation snapshot.

Evidence:

packages/coding-agent/src/core/context-budget-v2-planner.ts
packages/coding-agent/src/core/context-budget-v2-scoring.ts
packages/coding-agent/src/core/context-budget-v2-selection.ts
packages/coding-agent/test/context-budget-v2-knapsack-order.test.ts
packages/coding-agent/test/context-budget-selection-policy-version.test.ts
packages/coding-agent/test/context-budget-cache-disk.test.ts

Working tree: non-queued native xai and devin requests started through AgentSession.prompt() now derive a bounded automatic skill grant from live discovered descriptions after ordinary prompt-template expansion. The selector scores task text separately from camelCase-aware path-to-skill-name signals, excludes explicit-only skills, caps automatic matches at three, and adds headroom only under lexical or measured context pressure. AgentSession.prompt() merges the result with settings/SDK/bang selections only for that request. Queued steering/follow-up messages reuse the active run's system prompt and do not trigger another selection pass.

Evidence:

packages/coding-agent/src/core/active-skill-state.ts
packages/coding-agent/src/core/skill-selector.ts
packages/coding-agent/src/core/harness-skills.ts
packages/coding-agent/src/core/grok-harness.ts
packages/coding-agent/src/core/devin-harness.ts
packages/coding-agent/src/core/agent-session.ts
packages/coding-agent/test/grok-active-skills.test.ts
packages/coding-agent/test/devin-active-skills.test.ts
packages/coding-agent/test/skill-selector.property.test.ts

Working tree: context files now treat their global/local relevance baseline as a floor. Lexical overlap can raise that score but cannot demote standing instructions below the no-query baseline. Skills remain topic-scored because they are optional capabilities, not standing authority. The change is in scoreContextFileRelevance() with regression coverage in packages/coding-agent/test/context-budget-relevance.test.ts.

Working tree: the default compactor now strips model-generated managed-rule sections and deterministically carries explicit user-authored RULE/ INVARIANT/CONSTRAINT/MUST/NEVER/ALWAYS markers (plus explicit Korean markers) outside LLM rewriting. Rules are bounded to 64 × 1,000 characters, stored in additive non-hook compaction details, and covered by a five-round byte-preservation test. Natural-language classification, hook summaries, branch summaries, and cross-session memory remain outside this slice. Persisted rules are credential-redacted and source-bound to a user entry/line/digest; previous details are reused only when their canonical block matches the prior summary.

Evidence:

packages/coding-agent/src/core/compaction/knowledge-triage.ts
packages/coding-agent/test/compaction-knowledge-triage.test.ts
packages/coding-agent/test/compaction-summary-reasoning.test.ts

Reasoning routing

Status: Released / opt-in. /think auto uses the local, deterministic v4 router. It extracts bounded prompt features, classifies one of seven task classes, maps the class through TASK_CLASS_THINKING_LEVELS, applies lane steps, bounded bias/hint adjustments, and non-negative uncertainty escalation, then clamps the result to the selected model's supported levels.

The released extension-signal coefficients are active and bounded: multiTurnPrior=2, pressureBucket=1, and judgeVote=2. History and pressure are supplied by the main session; a judge vote affects only callers that provide one. A zero-score fallback cannot be hijacked by these signals.

The optional learning path is global-only and off by default. When enabled, a session loads one strictly validated RouterBiasSnapshot from the configured path or repository-scoped default, pins that snapshot or miss for the session, applies a -2..2 step bias, and writes only bounded feedback buckets. It never stores prompts, diffs, tool output, provider payloads, or repository paths.

Evidence:

packages/coding-agent/src/core/reasoning-router-v4.ts
packages/coding-agent/src/core/reasoning-router-resolver.ts
packages/coding-agent/src/core/reasoning-router-bias.ts
packages/coding-agent/test/suite/regressions/013-reasoning-router-v4-accuracy.test.ts
packages/coding-agent/test/suite/regressions/014-reasoning-router-v4-learning-wiring.test.ts
packages/coding-agent/test/suite/regressions/018-reasoning-router-v4-inert-weights.test.ts

Working tree: promotion evidence now credits a row only when repeated baseline and candidate classifier replays each agree. It also requires a frozen baseline and fails closed on insufficient or unstable replays. See reasoning-router-replay-stability.ts, reasoning-router-policy-ceiling.test.ts, and reasoning-router-replay-stability.test.ts.

Resource governance, lanes, and shards

The resource plane probes memory, workspace disk, V8 heap, and system CPU, then produces bounded admission caps.

Mechanism	Status	Authority
Prompt-time probe, admission decision, and journal	Released / default	Observe and record
`/resource` and `omk doctor resources`	Released / opt-in	Inspect current policy and probe state
`omk doctor resources --report`	Working tree	Aggregate bounded local admission evidence; never promotes mode
Per-run tool cap and governed heavy-process permits	Released / opt-in	Enforced in `adaptive` or `strict` mode
`launchSubagentLanes()`	Released / internal	No live child-dispatch consumer
Journaled Vitest/Jest/workspace/Go shard executor	Released / internal	No `autoShard` setting or session-command consumer

observe remains the default. Admission caps never raise configured caps. Corrupt shard journals are quarantined and block resume; completed shards may be skipped, but shard completion is only evidence and never a task verdict.

Evidence:

packages/coding-agent/src/core/resource-admission.ts
packages/coding-agent/src/core/resource-governor-settings.ts
packages/coding-agent/src/core/run-resource-lease.ts
packages/coding-agent/src/commands/resource-doctor-cli.ts
packages/coding-agent/src/core/resource-observation-report.ts
packages/coding-agent/test/resource-observation-report.test.ts
`packages/coding-agent/src/core/subagent-lane-launcher.ts
packages/coding-agent/src/core/workload-shard-executor.ts
packages/coding-agent/test/resource-admission.test.ts
packages/coding-agent/test/resource-doctor-cli.test.ts
packages/coding-agent/test/agent-session-resource-lease.test.ts
packages/coding-agent/test/agent-session-resource-permits.test.ts
packages/coding-agent/test/subagent-lane-launcher.test.ts
packages/coding-agent/test/workload-shard-executor.test.ts

Terminal settlement notifications

Status: Working tree. v0.97.0 released completion sound as opt-in and suppressed user aborts. The current working tree enables it by default on an interactive TTY and adds an onAbort outcome switch. Successful prompts retain the 5-second duration floor; failed and aborted/stopped outcomes notify immediately.

The sound consumes only prompt_settled, never intermediate agent_end. The current live path drains provider attempts, tools, and queued continuations; subagent work is awaited inside its tool call. The settlement reducer reserves direct child/shard counters, but no production signal call wires them yet, so future live lanes/shards must close that gap before activation. RPC, JSON, print mode, and CI remain silent. Playback uses fixed absolute executable/argv pairs, a minimal environment without inherited PATH or credentials, and a neutral temporary cwd. WSL uses BEL rather than resolving PowerShell through PATH. Playback is fire-and-forget and cannot change the prompt outcome.

Evidence:

packages/coding-agent/src/core/prompt-settlement.ts
packages/coding-agent/src/core/completion-sound.ts
packages/coding-agent/src/core/completion-sound-io.ts
packages/coding-agent/test/prompt-settlement.test.ts
packages/coding-agent/test/completion-sound.test.ts
packages/coding-agent/test/suite/agent-session-retry-events.test.ts

Evidence, decisions, and recovery

Status: Released / opt-in. The explicit omk-protocol API provides versioned, readonly record contracts and runtime parsers for tasks, attempts, observations, evaluations, waivers, and runtime decisions. evaluateTask() is pure. Among unwaived required claims, any violation yields fail; otherwise missing evidence yields inconclusive; otherwise the verdict is pass. A task with no required claims is inconclusive; explicit waivers remove their claims from verdict reduction. Advisory judging can choose among candidates that already passed; it cannot create evidence or change a semantic verdict.

The coding-agent adds evidence receipts, a replay ledger, workspace fingerprints, attempt journals, durable goals, seam checkpoints, session doctor, and bounded provider retry/failover. Digests detect mismatch; they do not prove runner honesty, OS isolation, freshness, or trusted authorship by themselves.

Evidence:

packages/protocol/src/evaluation.ts: evaluateTask
packages/protocol/src/decision.ts: reduceRuntimeDecision
packages/protocol/test/protocol.test.ts
packages/coding-agent/src/core/advisory-judge.ts
packages/coding-agent/test/advisory-judge.test.ts

Working tree (2026-09-05): the first-party createModelAdvisoryJudge() adapter rejects non-normal or missing completion metadata. Custom judges still supply raw score JSON and own that metadata boundary. The chooser checks cancellation around asynchronous work, reports top-score ties as caller-rank decisions, and retains submitted/eligible/excluded counts with comparison availability. This adds no completion calls or retries, automatic AgentSession/TUI call, or semantic-verdict authority. Details: Advisory selection integrity; governing spec: specs/021-advisory-selection-integrity/spec.md.

See Run Protocol and Durable Goals, Sessions, Provider Resilience, and Turn Metrics.

AdaptOrch and WPL boundary

Status: Released / opt-in. Published omk-adaptorch-wpl supplies typed packet state, client, adjudication, and verdict-projection primitives. Its loop.ts explicitly excludes end-to-end adaptorch_run dispatch, polling, request assembly, and persistence. The coding-agent has no production importer that turns those primitives into a default execution loop; the Correctness Wall remains an explicitly loaded example extension. The default-off AdaptOrch reasoning bridge currently returns no advisory hint.

Evidence:

packages/adaptorch-wpl/src/loop.ts
packages/adaptorch-wpl/test/loop.test.ts
packages/coding-agent/src/core/adaptorch-bridge.ts
packages/coding-agent/test/suite/regressions/011-reasoning-router-adaptorch-bridge.test.ts
packages/coding-agent/test/suite/regressions/012-reasoning-router-learning-adaptorch-activation.test.ts

See AdaptOrch Preview and Correctness Wall.

Repository understanding

Status: Working tree. The v0.97.0 release shipped policy/workflow only and still ignored /openwiki/ and contained neither the corpus nor its checker. Neither the release tag nor the current Git index tracks openwiki/ or .understand-anything/; both datasets are untracked or ignored advisory state. The current .last-update.json says interrupted at current HEAD.

scripts/check-openwiki.mjs validates entry pages, internal links, update-state shape, and global symbol-name presence in a source/test/script haystack. It does not validate prose or bind each symbol to a declared source path. A current-HEAD interrupted generation warns; a stale interrupted generation blocks. Source and tests outrank every generated page. There is no dedicated checker test in the current working tree; the executable checker is the available evidence.

This is not a release-grade trust gate: the current corpus is interrupted, symbol checks are global substrings rather than declared source bindings, and the generation workflow still needs an output allowlist plus pre-upload secret and private-path scans. Do not load the corpus as trusted context until those blockers close.

Direction

Status: Proposed. Owning specification: specs/015-runtime-algorithm-direction/spec.md. This section creates no runtime or test evidence; the internal mechanisms above do not satisfy these promotion gates. Apply them in order:

Measure before promoting authority. Run dated, same-model, same-provider, same-task comparisons before changing defaults or making leadership claims. The protocol in Turn Metrics is mandatory.
Reduce structure before adding mechanisms. Do not raise module-size, dependency-tree, or import-cycle baselines. The dependency-tree and import- cycle gates are working-tree changes. Split large session and interactive modules by ownership; do not mix movement-only refactors with behavior.
Promote live authority in stages. Collect resource observations before making adaptive the default. Wire subagent lanes before exposing automatic sharding. Keep automatic sharding opt-in and limited to known, semantically equivalent command families.
Protect standing context. The relevance floor and explicit-rule compaction slice are Worktree-only. Broader natural-language, hook, and branch triage still require evidence before expansion.
Design verified memory before implementing it. Spec 019 requires evidence-linked admission, Context Budget V2 data-only injection, source-span provenance, staleness, and memory-injection probes. Implementation remains blocked until fixtures and evaluation thresholds are preregistered.
Keep adaptation evidence-gated. Do not add online router learning while the current instrument cannot show gain after BIAS_STRONG_THRESHOLD=5. Reopen only through an advance spec that preregisters the real outcome- linked sample, sample size, minimum effect, confidence rule, and statistical test before collecting the promotion result.
Keep hosted advice separate from local authority. AdaptOrch remains a separate service and any bridge remains advisory; local deterministic gates own execution and completion.

Deliberate non-goals

Calling a level-barrier schedule an eager critical-path executor
Treating internal lane or shard code as a live feature
Auto-sharding arbitrary shell, deploy, publish, release, or migration commands
Treating model narration, reviewer opinion, or a digest alone as completion
Persisting unrestricted prompts or trajectories as learning memory
Claiming SOTA from tests, feature counts, self-scores, or roadmap projections

Verification map

Claim	Fast verification
Documentation links resolve	`npm run check:doc-links`
Specification governance holds	`npm run check:constitution`
Context selection behavior	focused context-budget tests named above
Router behavior and promotion ceiling	focused reasoning-router tests named above
Tool scheduling and timeout candidate	focused `packages/agent` scheduler/timeout tests
Resource and internal execution mechanisms	focused resource, lane, and shard tests named above

A green focused test proves only its declared behavior. Release readiness still requires the repository's full release gates.