Skip to main content
Glama
demeet2k

Athena MCP Server

by demeet2k
README.md
# ATHENA Canonical MCP v3.4 — AΩR × Collective V1–V15 × Operational Organism × Deployment.2

ATHENA is an executable Git/MCP developmental substrate. The current runtime composes canonical identity/versioning, typed JSPACE, SCALE, KC144/polycoordinates, AOR/Y1 developmental routing, Collective V1–V15 science/control layers, restart-safe state foundation, prompt/frontier and rehydration loops, Message Board/cohesion/party coordination, campaign/life/game organs, bounded symbolic/mythic/bionano kernels, Deployment.2 external-control planning, versioned release distribution, and exact-head qualification through a host-bound GitHub verifier.

Current executable coordinates:

- package: `athena-canonical-mcp 3.4.0`;
- live architecture: `ATHENA.RUNTIME.UNIFIED.11`;
- runtime root: one composed `Server`;
- Collective ladder: V1–V15;
- current Collective coordinate: `COLLECTIVE_CALIBRATED=<SR,XT,XD,CJ,AT,MD,L>`;
- inherited V14 coordinate: `COLLECTIVE_SYNTHESIS=<JB,SE,JE,DR,RP,AZ,MR,L>`;
- deployment frontier: `ATHENA.DEPLOYMENT.2`;
- authority: Y1 `athena_claim_*`;
- science-shadow namespace: `athena_discovery_claim_*`;
- governance: `SCHEMA.2 / SELFTEST.1 / SURFACE.2 / COMPOSITION.2 / PROMOTION.2`;
- trusted verifier: `GITHUB_PROMOTION_VERIFIER.1`;
- trusted qualification: `syntax ∧ unit ∧ critical-invariants ∧ smoke → promotion-qualification`.

## Constitutional braid

`AOR = WHAT is developmentally eligible`.

`COLLECTIVE = HOW bounded science/inference/control capacity is organized`.

`Y1 = canonical claim authority`.

`OPERATIONAL ORGANISM = prompt/frontier + rehydration/successor/handoff + Message Board/cohesion/party + campaign/life/game + symbolic/mythic/bionano organs`.

`DEPLOYMENT.2 = separately typed external-control planning / validation / canary / receipt verification`.

`PROMOTION.2 = caller-bound readiness + separately trusted qualification`.

Core firewalls:

`UNKNOWN != 0`

`CONSENSUS != EVIDENCE`

`PREDICTION / POSTERIOR / PLAN != OBSERVATION / TRUTH / EXECUTION`

`MODEL_GRAPH != CANONICAL_JSPACE_GRAPH`

`CALLER_ATTESTATION != TRUSTED_EXTERNAL_VERIFICATION`

`ATTESTED_READY != QUALIFIED`

`COLLECTIVE_CALIBRATED != DEPLOYMENT_AUTHORITY != COORDINATION_AUTHORITY`.

## Drive Wiki continuity candidate

The `ATHENA.WIKI.MEMORY.V1` extension imports version-bound local Wiki observations and recovers their claims, evidence, conflicts, handoffs and document text after a process restart. Five `athena_wiki_*` tools share the existing server and database. See [Wiki memory V1](docs/wiki-memory-v1.md) for the import protocol, provenance limits and tests.

The Git bridge can [stage a reviewed local draft](docs/wiki-draft-v1.md), [read a historical draft](docs/wiki-draft-review-v1.md), [query a complete committed Wiki](docs/wiki-committed-query-v1.md), and [open the exact bytes behind a source citation](docs/wiki-committed-source-v1.md). These operations preserve source identities and uncertainty. The shared [snapshot reader](docs/wiki-snapshot-read-v1.md) loads bounded committed evidence with two Git processes without checking it out.

## Runtime cycle

`HYDRATE → RECONRUN/OMEGA → MEMORY → SX → RAG → HUG → GAP → FIELD → MEASURE/CALIBRATE → Y1/AOR → COLLECTIVE(V1–V15) → AUTHORIZED EXECUTION → VERIFY → LEARN → SUCCESSOR → COMPLETE`.

CYCLE stops at typed missing prerequisites instead of fabricating execution, evidence, workers, tests or authority.

## Collective V1–V14 lineage

V1–V9 provide organization, memory, empirical learning, ecology, causal science, active discovery, dual control, finite belief and Gaussian/continuous inference. V10–V13 add bounded GP world models, adaptation, joint hypermodel uncertainty, FCI-lite/longitudinal causal estimation and robust resources. V14 adds finite joint factor belief, bootstrap structural ensembles, joint decision/information value, sequential DR policy value, robust policy geometry, decision-relative GP zoom and finite two-stage recourse.

Resource: `athena://collective/v14`.

V14 remains historical/currently callable science substrate; V15 does not silently rename it.

## V15 — calibrated continuous scientific control

### SR — externally witnessed structural reliability

`athena_structural_reliability_calibrate` fits a weighted monotone isotonic mapping from bootstrap structural support to externally labelled correctness. Diagnostic predictions are out-of-fold.

Identical support coordinates are aggregated before PAV; the final mapping uses an explicit right-continuous monotone step convention with endpoint extension.

`IDENTICAL_CALIBRATION_COORDINATE != MULTIPLE_FITTED_VALUES`.

`OUT_OF_FOLD_ISOTONIC_RELIABILITY != CAUSAL_GRAPH_POSTERIOR`.

`CALIBRATION_LABELS_REQUIRE_EXTERNAL_WITNESS`.

Calibration never writes JSPACE.

### XT — history-safe cross-fitted sequential TMLE

`athena_longitudinal_tmle_crossfit` is bounded to binary two-timepoint histories `X → A1 → L1 → A2 → Y`. Nuisance and targeting models are trained without each held-out evaluation fold.

The stage-2 pseudo-outcome preserves each row's observed `A1,L1` while intervening only on `A2`; stage-1 intervention/evaluation occurs later.

`STAGE2_PSEUDO_OUTCOME_PRESERVES_OBSERVED_A1_L1_BEFORE_STAGE1_INTERVENTION`.

Named treatment/intermediate/outcome fields cannot be smuggled into baseline covariates, and baseline values must be finite.

`CROSS_FITTED_TWO_TIMEPOINT_TMLE != GENERAL_LONGITUDINAL_TMLE_THEOREM`.

`CROSS_FITTING != IDENTIFICATION`.

Declared latent confounding fails closed.

### XD — history-safe cross-fitted sequential DR policy value

`athena_sequential_dr_policy_crossfit` evaluates deterministic two-timepoint policies with out-of-fold sequential AIPW scores.

Decision-time information sets are explicit:

`A1_POLICY_FEATURES = baseline`

`A2_POLICY_FEATURES = baseline + {A1,L1}`.

A stage-1 policy therefore cannot read `L1`, `A2`, or `Y`; a stage-2 policy cannot read `A2` or `Y`.

`DECISION_TIME_HISTORY != FULL_ROW_STATE`.

`A1_POLICY_USES_BASELINE_ONLY__A2_POLICY_USES_BASELINE_A1_L1_ONLY`.

`CROSS_FITTED_SEQUENTIAL_DR != GENERAL_OFF_POLICY_CAUSAL_VALUE`.

Policy value remains PLAN_ONLY.

### CJ — strict continuous Gaussian joint belief/control

`athena_joint_gaussian_update` implements the exact finite-dimensional update for `X~N(mu,Sigma)` and a declared linear Gaussian observation `y=h^T X+epsilon`.

Unknown observation/action coefficient keys are rejected rather than silently projected to zero; means, covariances and control parameters must be finite.

`UNKNOWN_COEFFICIENT != ZERO_COEFFICIENT`.

`NONFINITE_NUMERIC_STATE != MODEL_COORDINATE`.

`LINEAR_GAUSSIAN_UPDATE != GENERAL_CONTINUOUS_JOINT_BAYES`.

`athena_joint_gaussian_control` propagates linear action utilities through that Gaussian belief and retains expected utility, lower-tail Normal CVaR, cost and Pareto state.

`GAUSSIAN_LINEAR_CONTROL != GENERAL_BELIEF_MDP`.

### AT — local/global approximation-error transport

`athena_approx_error_transport` validates a caller-declared Lipschitz error envelope against supplied witness pairs.

The unrestricted mathematical envelope remains:

`e_global(x) <= min_i [e_i + L ||x-x_i||]`.

But V15 keeps three coordinates distinct:

- geometrically nearest witness;
- tightest global envelope witness;
- tightest radius-eligible local transport witness when a radius is declared.

A local radius certificate uses only eligible witnesses; the unrestricted global envelope is reported separately.

`GEOMETRIC_NEAREST_WITNESS != TIGHTEST_ERROR_ENVELOPE_WITNESS`.

`GLOBAL_ENVELOPE != RADIUS_ELIGIBLE_LOCAL_CERTIFICATE`.

`DECLARED_LIPSCHITZ_ERROR_ENVELOPE != EMPIRICAL_GLOBAL_ERROR_TRUTH`.

`TRANSPORT_CERTIFICATE_CONDITIONAL_ON_LIPSCHITZ_BOUND`.

### MD — strict finite-horizon rectangular TV-DRO

`athena_multistage_tv_dro_plan` solves exact backward induction for supplied finite states/actions and state-action rectangular total-variation ambiguity:

`V_t(s)=max_a[r(s,a)+gamma min_{q:TV(q,p_sa)<=rho} q^T V_{t+1}]`.

State/action identities, rewards and transitions must be finite and declared. Unknown successor/state coordinates fail closed instead of being ignored.

`UNKNOWN_STATE_COORDINATE != UNUSED_METADATA`.

`NONFINITE_TRANSITION != PROBABILITY_MODEL`.

Certificate:

`EXACT_DYNAMIC_PROGRAM_FOR_SUPPLIED_FINITE_RECTANGULAR_TV_AMBIGUITY_MODEL`.

`RECTANGULAR_TV_ROBUST_MDP != GENERAL_MULTISTAGE_DRO`.

Resource: `athena://collective/v15`.

Specs: `spec/COLLECTIVE_RUNTIME_V15.md`, `spec/ATHENA_UNIFIED_V15.md`, `spec/ARCHITECTURE_V15.md`, `spec/MIGRATION_V15.md`.

## V15 release-overlay holonomy

Ω15 discovered that one semantic runtime coordinate can exist as multiple import-time Python projections: module attributes, values/functions imported by value, copied resource lists, and derived URI sets.

The V15 overlay therefore synchronizes initialize/HTTP server identity, manifest builders, MAXDEV fallback, integrity resources, and final AOR development resources explicitly.

`RELEASE_ATTRIBUTE_UPDATE != IMPORTED_VALUE_SNAPSHOT_UPDATE`.

`MODULE_ATTRIBUTE_ADVANCE != IMPORTED_FUNCTION_SNAPSHOT_ADVANCE`.

`SOURCE_RESOURCE_ADVANCE != COPIED_RESOURCE_REGISTRY_ADVANCE`.

Tool manifest, `athena://runtime/unified-manifest`, and `athena://manifest` are regression-tested as one current release coordinate.

Release critical selectors are also checked against real repository test files:

`ZERO_TEST_SELECTION != PROOF`.

## Deployment.2 braid

Live `master` evolved Deployment.2 while Ω15 was in flight. Ω15 therefore uses a true two-parent Git braid and composes the initializer as:

`V14 → Deployment.2 → V15 release overlay`.

Deployment tools include `athena_deployment_manifest`, `athena_deployment_validate`, `athena_deployment_activation_plan`, `athena_deployment_assess_canary`, and `athena_deployment_verify_receipt`.

Deployment resources include `athena://deployment`, `athena://deployment/security`, `athena://deployment/rollout`, and `athena://deployment/evidence`.

Deployment is external-control state; planning or validating a deployment is not proof that production infrastructure changed.

## Trusted promotion topology

`athena_promotion_evaluate` remains caller-bound and can reach `ATTESTED_READY` from exact-head caller packets.

`athena_promotion_verify_github(git_head)` is the host-bound route. Repository, API root, Actions run, token, trusted app and required checks are not caller-selected MCP arguments. One coherent Actions suite must contain successful `{syntax, unit, critical-invariants, smoke}` checks for one exact head before `promotion-qualification` can mint a trusted receipt.

`CHECKS_FROM_DIFFERENT_SUITES_OR_RUNS != ONE_TRUSTED_QUALIFICATION`.

`VERIFIER_IMPLEMENTED != HEAD_QUALIFIED`.

## Release distribution

V3.2 and V3.3 manifests/notes/frozen regressions remain historical evidence for their exact bytes. V3.4 owns the current executable distribution workflow.

`OLD_RELEASE_RECEIPT != NEW_RELEASE_EVIDENCE`.

`HISTORICAL_PUBLICATION_AUTHORITY != CURRENT_RELEASE_IDENTITY`.

Release qualification:

`syntax ∧ unit ∧ critical-invariants ∧ smoke → promotion-qualification → package-readiness`.

The V3.4 critical lane explicitly executes hardened V15 calibration geometry, adversarial input/temporal/numeric membranes, surface holonomy, Deployment.2 composition and trust boundaries. A distribution receipt certifies exact repository/package/distribution state. It is not empirical truth, causal proof, treatment authorization, production deployment, Y1 authority or GitHub administrative hardening.

## Historical architecture

`ARCHITECTURE.md` / `MIGRATION.md` preserve the V13/UNIFIED.9 transition. V14 history remains in `spec/ARCHITECTURE_V14.md` and `spec/MIGRATION_V14.md`. Current V15 composition/migration are versioned separately, preserving the actual succession rather than rewriting prior evidence.

## Run

The [Windows continuity contract](spec/WINDOWS_CONTINUITY_V1.md) describes
byte-preserving prompt writes, portable receipt paths and test-process handling.
Run the complete suite with `python -m unittest discover -s tests -v` on either
platform; CI includes both Linux and Windows discovery runs.

The Wiki memory tools preserve source-bound Drive registry and document
observations across sessions. `athena_git_wiki_compile` invokes the configured
semantic repository's fixed Wiki compiler at an explicit clean commit and
returns its original receipts and proposed changes. See
[Git Wiki compiler through MCP](spec/WIKI_GIT_COMPILER_V1.md) for the data,
execution and application boundaries.

`athena_wiki_git_ingest` connects a specific imported observation to that
compiler: it preserves an immutable source carrier, retains committed Wiki
pages, and compiles INGEST, REINDEX and LINT into one guarded proposal. See
[Drive observation ingestion](spec/WIKI_GIT_INGEST_V1.md).

`athena_wiki_git_stage` materializes a reviewed ingestion plan as a verified
local draft commit, with fresh committed lint and create-only ref publication.
See [local Wiki review commits](docs/wiki-draft-v1.md).

`athena_wiki_git_review` discovers local drafts and reads a pinned draft's
source carrier after the checkout advances, without a Wiki database or
repository-code execution. See [read committed drafts](docs/wiki-draft-review-v1.md).

`athena_wiki_git_query` reads complete evidence directly from a pinned Git Wiki
commit and queries it with the configured trusted compiler, without requiring
caller-built collections or changing the checkout. See
[query committed Wiki evidence](docs/wiki-committed-query-v1.md).

`python -m athena_mcp --db ./state/athena.db`

Package: `athena-canonical-mcp 3.4.0`  
MCP protocol revision: `2025-11-25`.

TDQS

C2.5/5.0

Scored across 336 tools

Disambiguation2/5

Many tools share the same action suffix (observe, predict, evsi, get, recent, replay) across different model families; for example, there are at least ten *_observe tools and several decision-EVSI variants whose boundaries are only clear from dense descriptions. While domain prefixes help, an agent would frequently need to read lengthy descriptions to avoid mis-selecting between similar causal, GP, Bayesian, and belief variants.

Naming Consistency4/5

The vast majority follow a consistent `athena_<domain>_<action>` snake_case pattern, and the domain prefix provides predictable grouping. However, a minority use bare verb names (hydrate, orchestrate, register, search) or verb-first names (add_edge, apply_transform, reconstruct_state), so the convention is not perfectly uniform.

Tool Count1/5

336 tools is an extreme mismatch for an agent-facing surface. Even with broad domain ambitions, this overwhelms context windows and tool-selection; it should be split into multiple focused servers or drastically consolidated.

Completeness3/5

The surface covers an impressively wide range of domains and many resources have get/recent/replay/state lifecycles. However, several resources lack obvious lifecycle operations (e.g., party leave/disband, branch create/delete, claim revoke/update), and the many model-specific variants do not add distinct capabilities, leaving notable gaps.

Maintenance

ActivityMaintained
ResponsivenessResponsive