Common agents and methods
P A skill is a complete procedure that can use LLM judgment and optional executable helpers. A tool is runnable source and/or a binary inside that procedure. Connector calls, state transitions, agent calls and permission checks are runtime access, not extra learned skills or imaginary binaries. Only the methods and segments selected for the actual case run.
Module-owned definitions
- Support Root — agent and method subpage.
- Data & Manipulation — agent and method subpage.
- Configuration — agent and method subpage.
- Customer-specific development — agent and method subpage.
- Assessment & Offers — agent and method subpage.
P This catalogue defines only shared capabilities. Domain-specific content stays with the module that owns it. A common definition is reused through separately scoped tasks, not shared transcripts or authority.
Agents
common.communicator — Closure and customer reply
P Function. Draft and communicate the scoped customer/owner brief: clarification, verified result, assessment questions, offer terms or decision. Preserve audience, approved facts and exact terms; sends and case closure remain controlled by the root.
| Contract | Content |
|---|---|
| Input brief | Current request/clarification, assessment, offer, decision or verified-result paper, as required by this stage Audience, register/language, source values/conditions, open questions and exact permitted send scope |
| Output brief | Audience-specific draft preserving source facts and terms, or a scoped request for missing information Approved-send evidence and communication disposition returned to the accountable root |
| Scope and restrictions | One scoped task instance per selected responsibility; a reused profile never shares another case transcript, grants or mutable task state. Boot from the own domain index, task request, pinned profile and confirmed parent plan. Select relevant articles, skills, tools and helpers on demand. No DB tools and no authority to mark an unverified effect resolved. Root still owns closure; delivering a reply is not acceptance. No invented price, commitment or delivery time; no Core Development implementation pipeline follows acceptance. Optional on the first stage of every lane: called when that stage cannot start for want of material or authority, or when the incident cannot wait. It asks or alerts, the case waits with actor, trigger and age, and the same stage resumes on the answer. It never owns the case, its clocks or its route. |
| Memory / ownership | support (communication rules) |
| Role instructions | Keep origin, intent, route, parent-confirmed plans, gates, accepted handovers, waits and obligations coherent. |
| Runtime / connector access | case.read / papers / task state connector.channel.send (gated) |
| Eval set | support.closure — declared by the owning module manifest or platform-profiles.json; green before publication (Agents § 0.1) |
| Experience strategy | owned-domain; retrieval and index contract |
| Runtime profile identity | support.communicator |
Methods used. common.skill.communication.
Executable helper scope. common.tool.aisa-jira, common.tool.aisa-mail, common.tool.aisa-hdesk. These are used through the selected method, not an extra workflow.
Source basis. Components, customer reply on point, ticket closure via pending.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
Former reference support.communicator resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.communicator.
common.experience-curator — Experience curator
P Function. Receive verified case outcomes from all modules, preserve actual outcomes, and coordinate post-result learning and scheduled consolidation.
| Contract | Content |
|---|---|
| Input brief | Verified reports and source-request identity, reuse/failure observations, domain-owner proposals |
| Output brief | Sanitised experience entry, owner-addressed capability changes and reviewed active index revisions |
| Scope and restrictions | One scoped task instance per selected responsibility; a reused profile never shares another case transcript, grants or mutable task state. Boot from the own domain index, task request, pinned profile and confirmed parent plan. Select relevant articles, skills, tools and helpers on demand. Post-result work is outside the active delivery process; a required repair to a failed skill remains an explicit case obligation. No estate connectors or foreign-domain activation. Shared capability identity does not widen the requesting case data class, model route, environment, evidence visibility or permitted effects. |
| Memory / ownership | support/experience (owned through runtime support.curator); other domains by proposal |
| Runtime / connector access | runtime.project-owned-knowledge experience.query (strategy-selected) runtime.validate-package-binding |
| Eval set | support.curator — declared by the owning module manifest or platform-profiles.json; green before publication (Agents § 0.1) |
| Experience strategy | owned-domain; retrieval and index contract |
| Runtime profile identity | support.curator |
Methods used. common.skill.retained-learning, common.skill.experience.
Selected segments. common.skill.experience: owned-domain strategy; current case/tenant and version filters
Source basis. Agents Memory, Components, Stage Precipitation.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
common.db-extractor — Stored-source extractor
P Function. Capture immutable per-environment source/signatures and dependencies without making code changes. A generic read-only extraction any module calls: Customer-specific development for its planning baseline, Configuration and Support when a stored procedure is the evidence.
| Contract | Content |
|---|---|
| Input brief | Scoped environment and stored-code object list |
| Output brief | Spec/body snapshot, signature, status/errors and caller/dependency evidence |
| Scope and restrictions | One scoped task instance per selected responsibility; a reused profile never shares another case transcript, grants or mutable task state. Boot from the own domain index, task request, pinned profile and confirmed parent plan. Select relevant articles, skills, tools and helpers on demand. Read only. A focused one-tool extractor is intentional; it is not a stored-code implementer or a business-intent prober. Shared capability identity does not widen the requesting case data class, model route, environment, evidence visibility or permitted effects. |
| Memory / ownership | source/db-plsql (read) |
| Runtime / connector access | connector.oracle.source-and-signature |
| Eval set | source.planning — declared by the owning module manifest or platform-profiles.json; green before publication (Agents § 0.1) |
| Experience strategy | source-change; retrieval and index contract |
| Runtime profile identity | source.db_extractor |
Source basis. Components, Stage Planning.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
Former reference source.db_extractor resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.db-extractor.
common.compactor — Context compactor
P Function. Condense an existing task transcript into a resumable evidence-linked summary without changing durable state or decisions.
| Contract | Content |
|---|---|
| Input brief | Scoped transcript and result references |
| Output brief | Decisions, sourced facts, open items and result refs in a summary artefact |
| Scope and restrictions | One scoped task instance per selected responsibility; a reused profile never shares another case transcript, grants or mutable task state. Boot from the own domain index, task request, pinned profile and confirmed parent plan. Select relevant articles, skills, tools and helpers on demand. No connector/tools or shared-memory writes. Summary cannot replace original papers or ledger truth. Shared capability identity does not widen the requesting case data class, model route, environment, evidence visibility or permitted effects. |
| Memory / ownership | No memory ownership |
| Eval set | platform.compactor — declared by the owning module manifest or platform-profiles.json; green before publication (Agents § 0.1) |
| Experience strategy | owned-domain; retrieval and index contract |
| Runtime profile identity | platform.compactor |
Source basis. Agents, Agent Runtime.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
common.summarizer — Knowledge-node summarizer
P Function. Read a missing source/domain surface and draft the smallest useful indexed node for its owner.
| Contract | Content |
|---|---|
| Input brief | Named missing domain, source/guidance/build/test/history scope and owner |
| Output brief | Case-local/proposed node with source links, build facts, traps and open gaps |
| Scope and restrictions | One scoped task instance per selected responsibility; a reused profile never shares another case transcript, grants or mutable task state. Boot from the own domain index, task request, pinned profile and confirmed parent plan. Select relevant articles, skills, tools and helpers on demand. No foreign-domain activation or invented source facts; not a new permanent profile per repository. Shared capability identity does not widen the requesting case data class, model route, environment, evidence visibility or permitted effects. |
| Memory / ownership | Target parent domain; new node remains a proposal until owner confirmation |
| Runtime / connector access | connector.git.read runtime.project-owned-knowledge |
| Eval set | platform.summarizer — declared by the owning module manifest or platform-profiles.json; green before publication (Agents § 0.1) |
| Experience strategy | owned-domain; retrieval and index contract |
| Runtime profile identity | platform.summarizer |
Source basis. Agents, Agents Memory.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
common.grader — Blind evaluation grader
P Function. Compare candidate and reference artefacts against a declared rubric without provenance or author reasoning.
| Contract | Content |
|---|---|
| Input brief | Two sanitised artefacts and rubric only |
| Output brief | Score per rubric dimension, evidence quotes and Unknown where undecidable |
| Scope and restrictions | One scoped task instance per selected responsibility; a reused profile never shares another case transcript, grants or mutable task state. Boot from the own domain index, task request, pinned profile and confirmed parent plan. Select relevant articles, skills, tools and helpers on demand. No provenance, author identity/reasoning, connectors or self-publication. Kept separate from evidence-provenance-preserving validators. Shared capability identity does not widen the requesting case data class, model route, environment, evidence visibility or permitted effects. |
| Memory / ownership | No memory ownership |
| Eval set | platform.grader — declared by the owning module manifest or platform-profiles.json; green before publication (Agents § 0.1) |
| Experience strategy | owned-domain; retrieval and index contract |
| Runtime profile identity | platform.grader |
Source basis. Agents, Agent Framework.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
Domain procedures
common.skill.experience — Retrieve and judge experience for this task
P Function. Choose a retrieval strategy from the requesting agent’s domain and decision: classify a request, distinguish a mechanism, interpret a product change, plan a source change, or estimate effort. Search is one step; current applicability and evidence limits are LLM judgments.
| Contract | Content |
|---|---|
| Input brief | Purpose and requesting profile/domain; current tenant/case permissions; symptom or objective; environment/product/source versions; current index watermark |
| Output brief | Candidate case/article/skill references with owner/version and why selected; current applicability checks; rejected/stale matches; uncovered questions or investigation request |
| Scope and restrictions | Root classification reads a cross-domain routing/experience index and authorised open-case metadata, not every full memory or customer case. Domain agents start with their own methods and verified outcome patterns; broaden deliberately and record the strategy. Similar incidents are not duplicate requests; deterministic source identity and current case registry decide delivery deduplication. A retrieval miss from a stale index is not proof that experience does not exist. No old permission, price, tariff cell or production proof is carried into the new case. |
| Runtime / connector access | case-registry query owned-domain / routing-index query scoped source refresh through connector when required |
Source basis. Stage Solution Take, Agents Memory, myr policy missing last row pattern.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
Former reference common.skill.applicability resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.skill.experience.
common.skill.retained-learning — Retain and improve a capability
P Function. Convert verified outcomes and failed approaches into sanitised experience and an owner-reviewed memory/skill/tool change when useful.
| Contract | Content |
|---|---|
| Input brief | Case report, actual reuse/failure observations, repeated mechanical work and source identities |
| Output brief | Experience entry plus coherent optional article/skill/tool/test/index proposal; next-request applicability checks |
| Scope and restrictions | Selected for the current task when its applicability and published version fit; not loaded or executed merely because the stage exists. The procedure adds no target authority; every effect, including effectful verification, retains its own gate and journal. No one-output-file limit: source, tests, procedure and article may form one learning change. No automatic promotion from one success. Shared capability identity does not widen the requesting case data class, model route, environment, evidence visibility or permitted effects. |
| Runtime / connector access | runtime.project-owned-knowledge experience.query (strategy-selected) runtime.validate-package-binding |
Source basis. Agents Memory, Stage Precipitation.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
common.skill.communication — Prepare and deliver the outcome
P Function. Draft the answer to the actual request, with customer-safe and internal technical projections and the correct channel/acceptance state.
| Contract | Content |
|---|---|
| Input brief | Current request/clarification, assessment, offer, decision or verified-result paper, as required by this stage Audience, register/language, source values/conditions, open questions and exact permitted send scope |
| Output brief | Audience-specific draft preserving source facts and terms, or a scoped request for missing information Approved-send evidence and communication disposition returned to the accountable root |
| Scope and restrictions | Selected for the current task when its applicability and published version fit; not loaded or executed merely because the stage exists. The procedure adds no target authority; every effect, including effectful verification, retains its own gate and journal. Only the controlling root presents unresolved human authority; drafts omit raw personal data and unsupported completion claims. |
| Runtime / connector access | case.read / papers / task state connector.channel.send (gated) |
Optional executable helpers. common.tool.aisa-jira, common.tool.aisa-mail, common.tool.aisa-hdesk. Helper availability and role scope are checked separately.
Source basis. customer reply on point, ticket closure via pending, Components, Agents Memory.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
Former reference support.skill.communication resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.skill.communication.
common.db-prober — Measured database probe
P Function. Measure requested schema constraints and observed usage independently of the business outcome being proposed. The agent that runs it names the columns, objects and environment and is told nothing about the value it hopes for; the procedure returns what is permitted and what is used, kept separate.
| Contract | Content |
|---|---|
| Input brief | Named columns/objects and environment with narrowed read scope; no desired value or product intent |
| Output brief | Constraint/type and measured distributions, null/product/row counts and timestamp; permitted and used kept separate |
| Scope and restrictions | Selected for the current task when its applicability and published version fit; not loaded or executed merely because the stage exists. No product specification/intent, no recommendations and no writes. Reused by any module only for this same intent-blind measurement contract. Shared capability identity does not widen the requesting case data class, model route, environment, evidence visibility or permitted effects. |
| Memory / ownership | configuration (schema-measurement rules); connectors/oracle |
| Role instructions | Measure schema facts and distributions without seeing the caller desired outcome or selecting a business value. |
| Runtime / connector access | connector.schema.measure connector.oracle.read |
Source basis. Components, Stage Normalization, skill.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
common.effect-applier — Effect application — check the packet, then land it on the systems it touches
P Function. The procedure the owning stage agent runs to make an accepted result real on the systems it touches. First the four checks every effect packet passes before anything is asked of a target — exactness (every value bound, every quantity a number, no open predicate), boundedness (the affected set re-derived from the target equals the claim), reversibility (the revert exists, is derived, and restores exactly what the write changes), scope (every operation lies inside the grant and instantiates the approved shape). Then the landing, which needs the knowledge of those systems case by case: which objects take the system's own configuration endpoint and which take a parameterised statement, the order the calls must go in, the conventions that decide whether a write lands at all, what each failure mode actually means, and what proves the effect is there. Where a target exposes no operation it coordinates the assigned action instead; it journals what actually landed and reconciles anything unknown before a retry.
| Contract | Content |
|---|---|
| Input brief | The accepted result and its effect packet: statements or calls with bound parameters, expected counts, effect class, declared assertions and revert, plus the approved write shape and its version — never the author's transcript, reasoning or identity The case's grants, the answered gate decisions, and each target's connector row: its declared operations, its medium — endpoint · statement · publication sequence · user interface · person — its evidence, recovery and idempotency shape |
| Output brief | Per-check verdict bound to the packet hash: a pass, or the failing item with its evidence and no repair The ordered application across every target the result touches, and the durable intent/result journal: the operations actually run or the assignments actually raised, the measured outcome, proof references and landed, partial or unknown effect state |
| Scope and restrictions | Selected for the current task when its applicability and published version fit; not loaded or executed merely because the stage exists. The procedure adds no target authority; every effect retains its own gate and journal. The agent running it is never the author of the result it applies: it receives the packet, the approved shape and the grants — not the producing agent's transcript, reasoning or identity — and the procedure never repairs a packet that fails a check: it names the failing item and returns it. It asks; the runtime decides. Grants, gate decisions, shapes and ceilings are enforced by the authority service and the connector scopes, not by this profile's good behaviour: an operation outside them is refused before contact. Nothing about how an effect is made is shared between targets. Each connector row declares the medium — the system's own configuration endpoint, a parameterised statement template, an ordered publication sequence, a user-interface action, or an assigned person — with the operations, evidence, recovery and idempotency shape that belong to it. Two behaviours, one contract. Where the target declares an operation it causes the effect and reads that operation's evidence. Where it declares none it assigns: actor, rendered instruction, deadline, and the case parks. Nothing is simulated. An assigned action closes on a measurement of the target, never on the actor's report. What a person says they did is an input to the check, not the check. Shared capability identity does not widen the requesting case data class, model route, environment, evidence visibility or permitted effects. |
| Memory / ownership | platform/effects — write envelopes, connector conventions and failure decodes per system |
| Role instructions | Re-derive the specific claim from objective, evidence and provenance; return the evidence separating agreement from disagreement. |
| Runtime / connector access | runtime.validate-contract connector.oracle.read runtime.execute-granted-operation connector.what-landed |
Source basis. Agents, Gating, Connectors.Abstractions, Agent Runtime.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
Former reference common.write-auditor resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.effect-applier.
Former reference common.write-executor resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.effect-applier.
Former reference platform.write_auditor resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.effect-applier.
Former reference platform.write_executor resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.effect-applier.
Former reference common.skill.effect-application resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.effect-applier.
common.skill.validation — Validate independently: the result against the human-supplied scope
P Function. The isolation contract every domain validator applies with its own rubric: receive the objective and the human-supplied scope — the accepted specification, the approved plan or brief — the evidence with provenance and the exact artefact under review; never the author's transcript, identity or reasoning; re-derive the claim and return agree, disagree or unknown with the evidence that separates the alternatives. The producer verifies and tests its own work; the validator validates that work against the scope a person supplied. Common is the discipline; what is validated is always the domain's.
| Contract | Content |
|---|---|
| Input brief | Claim or objective, permitted evidence with provenance, the exact artefact revision and the domain rubric |
| Output brief | Evidence-backed agree / disagree / unknown verdict; a disagreement returns what would separate the two readings |
| Scope and restrictions | Selected for the current task when its applicability and published version fit; not loaded or executed merely because the stage exists. The procedure adds no target authority; every effect, including effectful verification, retains its own gate and journal. Not a runnable agent: the mechanism validator, the solution and outcome validators, the four configuration domain validators, the integration validator and the source validator are the agents; each binds this method to its domain rubric. Distinct from the effect applier, which checks an operation packet, not reasoning. |
| Runtime / connector access | runtime.run-declared-assertions compare current scoped evidence |
Source basis. Agents.
P Status. planned; seeded from the cited architecture and AISA workspace evidence; no published v2 capability is claimed
Former reference common.verifier resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.skill.validation.
Former reference common.skill.verification resolves to this responsibility. Existing task/profile revisions are not rewritten; new work uses common.skill.validation.
Executable helpers used by procedures
common.tool.aisa-blob — aisa-blob.exe — stored document extractor
P Function. Stream a BLOB column — a stored policy document, bill or attachment — from IPAL or INSIS into a file through the cached SQLcl connection metadata, in one call. Read-only evidence retrieval; it renders nothing and never produces a new document.
| Contract | Content |
|---|---|
| Input brief | Connection name, a BLOB-projecting SELECT or an SR_DOCS id, output path |
| Output brief | The stored bytes as a file plus the row identity used; nothing is written to the database |
| Scope and restrictions | Read-only; a retrieved document is evidence of what was stored, not of what a fresh render would produce. Credentials come from the cached SQLcl connection store; no secret enters a prompt or an evidence record. |
Executable evidence. Executable present; v2 package unregistered.
- Source: blob-cli.csproj.
- Binary: aisa-blob.exe.
Invocation: aisa-blob get-doc --connection NAME --sr-doc-id N --output FILE.
Source basis. aisa-blob README.
P Status. Executable present; v2 package unregistered.
common.tool.aisa-jira — aisa-jira.exe
P Function. Use the aisa-jira CLI source/executable as a connector implementation, for its documented commands and bounded argument set. The adapter, not the binary name, enforces the case’s read/send/change authority.
| Contract | Content |
|---|---|
| Input brief | Documented CLI command and scoped arguments; credential resolution stays with the adapter |
| Output brief | Actual structured command result/source record and evidence reference |
| Scope and restrictions | Read and mutating commands are distinct capabilities; a shared executable never confers permission to send or modify. No secrets in prompts, arguments logged as evidence or returned model context. |
Executable evidence. Executable present; v2 package unregistered.
- Source: aisa-jira.csproj.
- Binary: aisa-jira.exe.
Source basis. aisa-jira CLI documentation.
P Status. Executable present; v2 package unregistered.
common.tool.aisa-mail — aisa-mail.exe
P Function. Use the aisa-mail CLI source/executable as a connector implementation, for its documented commands and bounded argument set. The adapter, not the binary name, enforces the case’s read/send/change authority.
| Contract | Content |
|---|---|
| Input brief | Documented CLI command and scoped arguments; credential resolution stays with the adapter |
| Output brief | Actual structured command result/source record and evidence reference |
| Scope and restrictions | Read and mutating commands are distinct capabilities; a shared executable never confers permission to send or modify. No secrets in prompts, arguments logged as evidence or returned model context. |
Executable evidence. Executable present; v2 package unregistered.
- Source: aisa-mail.csproj.
- Binary: aisa-mail.exe.
Source basis. aisa-mail CLI documentation.
P Status. Executable present; v2 package unregistered.
common.tool.aisa-hdesk — aisa-hdesk CLI source — build required
P Function. Use the aisa-hdesk CLI source/executable as a connector implementation, for its documented commands and bounded argument set. The adapter, not the binary name, enforces the case’s read/send/change authority.
| Contract | Content |
|---|---|
| Input brief | Documented CLI command and scoped arguments; credential resolution stays with the adapter |
| Output brief | Actual structured command result/source record and evidence reference |
| Scope and restrictions | Read and mutating commands are distinct capabilities; a shared executable never confers permission to send or modify. No secrets in prompts, arguments logged as evidence or returned model context. |
Executable evidence. Source project exists; binary build is not present in this checkout.
- Source: aisa-hdesk.csproj.
Source basis. aisa-hdesk CLI documentation.
P Status. Source project exists; binary build is not present in this checkout.