Aller au contenu principal

Architecture: Agent prompt assembly

Status: Draft Last updated: 2026-05-25 Audience: Engineers wiring AgentRunService.execute() and the assistants reasoning about how a user-defined Agent's identity, memory, and constraints get into an AI call.

1. Purpose​

When a user-defined Agent runs (heartbeat tick, task execution, or chat reply), the platform must assemble a coherent system message + user-message stack that:

  1. Carries the Agent's identity (SOUL.md) and role (AGENTS.md).
  2. Tells the Agent what to do right now — heartbeat directive (HEARTBEAT.md), task description, or chat thread context.
  3. Constrains it via permissions (TOOLS.md and the agents.permissions JSON column).
  4. Augments it with skills (resolved per the hierarchy in agents-skills-tasks.md §8).
  5. Provides memory (recent runs, recent activity, scope state) without blowing the token budget.
  6. Honors the Work / Mission / Idea WorkAdvancedPrompts (existing per-scope prompt overrides) when the run is scoped to one.

The platform's existing AiFacadeService (packages/agent/src/facades/ai.facade.ts) does not assemble system messages today — pipeline plugins do that internally. The new AgentRunService.execute() will assemble the message itself and pass the final composed payload to AiFacadeService.createChatCompletion() / createStreamingChatCompletion() / askJson() via the existing messages parameter.

This doc specifies the canonical assembly order so different Agent triggers (heartbeat / task / chat) produce comparable prompts.

2. Assembly order — system message​

The system message is a concatenation of named segments. Lower index = earlier in the message = higher priority for the model's attention.

#SegmentSourceBudget (tokens)Required?
1Identity (SOUL.md)Agent's SOUL.mdfull✓
2Role (AGENTS.md)Agent's AGENTS.mdfull✓
3Capabilitiesagents.capabilities TEXT columnfulloptional
4Operating loopAgent's HEARTBEAT.md (heartbeat run only) OR a per-trigger preamble (task / chat)full✓
5Tools the Agent may callAgent's TOOLS.md filtered by permissions.canCallExternalTools etc.up to 1500✓
6Active SkillsSkills resolved by SkillBindingRepository.resolveActive (Skills spec §3.3). Progressive disclosure.maxSkillContextTokens (default 4000)optional
7Scope WorkAdvancedPromptsFor Work-scoped Agents, the relevant WorkAdvancedPrompts column (work-advanced-prompts.entity.ts)fulloptional
8Scope contextMission description / Idea description / Work initial_prompt — depends on scopeup to 800✓
9Memory: recent activityLast N=20 activity_log rows for this Agent's scope (filtered to events the Agent should "know about")up to 1200optional
10Memory: last N runsCompact JSON of the Agent's last 5 agent_runs summariesup to 800optional
11Output contractStrict JSON-schema reminder when caller used askJson(); otherwise free-form150depends

Total budget target: ≤ 12 000 input tokens in the system message. If a segment exceeds its budget, it is truncated tail-first (newest preserved, oldest cut). Truncation events emit an AgentRunLog row at level=WARN, step='prompt-assembly'.

Order rationale:

  • Identity + Role first — the model anchors on "who am I and what do I do" before "how do I do it."
  • Tools before Skills — Skills may reference tools; the tool list must be parsed first.
  • Skills before scope context — Skills should be domain-agnostic enough to apply across scopes.
  • Memory after instructions — memory tunes; instructions shape.
  • Output contract last — closest to the model's response, hardest to forget.

2.1 Per-trigger preamble (segment 4)​

For non-heartbeat triggers, segment 4 is replaced by a static preamble instead of HEARTBEAT.md:

Task execution preamble:

You are working on a specific Task assigned to you. The Task body
follows. Your output should advance the Task — make progress, ask a
clarifying question in the Task chat, transition the Task status,
or escalate by creating a sub-task. Do NOT take actions outside the
scope of this Task.

Chat reply preamble:

You were mentioned in a Task chat thread. Read the recent messages,
then post a single reply. Do NOT transition the Task status from a
chat reply — use the transition tool only when explicitly asked.
Keep the reply focused on the chat question.

3. User message​

Composed of:

#Content
1The immediate input — for heartbeat, the literal string "What's the next action you should take? Choose ONE."; for task, the Task description; for chat, the new chat message body.
2The conversation context — for task, the most recent 20 task_chat_messages of that task (newest last); for chat, the same. For heartbeat, omitted.
3Attachments / mentions / KB references rendered inline as fenced blocks the model can read. KB references injected as their instructionsMd body.

4. Memory model​

4.1 Short-term (within a run)​

The model sees segments 9-10 (recent activity + recent runs) as part of every run. This is the platform's "short-term memory" — last few hours / runs.

If the run does multiple tool-loop iterations (read a KB doc, then write a file, then post a chat), short-term memory accumulates inside the run as standard tool-loop messages. No additional storage; this lives in the in-flight LangChain conversation.

4.2 Long-term (across runs)​

The Agent's SOUL.md / AGENTS.md / HEARTBEAT.md / TOOLS.md are the long-term memory. They are durable, intentional, and editable.

Crucially: long-term memory is not automatically updated. The Agent has to actively decide "this is important enough to remember forever" and use the editAgentFile tool to append a paragraph. This keeps the model from polluting its identity files with conversational noise. See Agent self-editing rules below.

For passive history (last 30 days of activity), the model reads activity_log on demand via a getActivity({since, limit}) tool. Not injected by default — costs tokens.

4.3 Cross-Agent memory​

Agents cannot read each other's MD files. Cross-Agent knowledge flows through:

  • Tasks — assigning, commenting, mentioning.
  • Mission/Work KB — any Agent in scope can read KB documents.
  • Activity log — getActivity() includes other Agents' run summaries (but not their prompts).

This is intentional separation of identity (private) from work product (shared).

5. Token budgeting​

The dispatcher's pre-flight check (BudgetGuardService.checkBudget) estimates cost from an upper-bound token count: assembled system message + user message + 1500 reserved for the model's response. If the estimate would exceed remaining budget AND allowOverage = false, the run is short-circuited to failed with errorMessage='Budget exceeded'.

Actual usage is captured post-call by the existing PluginUsageEvent write inside AiFacadeService (research report Topic 1), with the new agentId and optional taskId columns set.

6. Conflict resolution​

6.1 Between Agent's MD files​

If SOUL.md says "be terse" and AGENTS.md says "be thorough" — the model is left to reconcile. We do not pre-resolve. The model's reasoning surfaces in its response and the user can iterate the files. The platform's only guarantee is order — SOUL is first, AGENTS is second, so identity has precedence over role description by attention-position bias.

6.2 Between MD files and user task​

User-task input wins. The Agent's HEARTBEAT.md may say "don't write code"; a Task with description: "write the migration" overrides for that Task. Modeled the same as a human employee: written role + ad-hoc assignment.

6.3 Between Agent's permissions and Skill's allowed-tools​

Permissions ALWAYS win. If permissions.canCommitToRepo = false, a skill whose allowed-tools includes git is rendered (the model sees it) but the commitToRepo tool returns a structured error when invoked. The skill body becomes context; not a back-door to the tool.

6.4 Between WorkAdvancedPrompts and Agent prompts​

WorkAdvancedPrompts is the per-Work override of the Work Generator's prompts. For Work-scoped Agents that are executing a Generator-relevant stage (e.g. an Agent specialized in "categorization"), the relevant WorkAdvancedPrompts column is injected as segment 7.

For Agents NOT executing a generator stage (e.g. the CEO doing strategy), segment 7 is skipped.

The decision is made by AgentRunService.execute() from the calling context: if the run was triggered by a Task whose body references a generator stage, fetch that column; otherwise skip.

7. Self-editing rules​

When the Agent calls editAgentFile({name, body}):

  1. Scope check — name ∈ {'SOUL.md', 'AGENTS.md', 'HEARTBEAT.md', 'TOOLS.md', 'agent.yml'} for the agent's own subtree only.
  2. Permission check — permissions.canEditAgentFiles = true.
  3. Secret scan — body must not match the secret-scan regex (sk-, xoxb-, AKIA…, ghp_…, glpat-…).
  4. Size cap — body ≤ 64 KB per file.
  5. Frequency cap — at most 1 edit per file per run (prevents tool-loop hammering).
  6. Hash check — if etag is supplied in the tool call, must match agents.contentHash (optimistic concurrency).
  7. Commit / inline write — for Mission/Work-scoped Agents, GitFacadeService.commit() with message chore(agent/<slug>): self-edit <file> via run <runId>; for Tenant-scoped Agents, write directly to the agents.{soulMd|agentsMd|heartbeatMd|toolsMd|agentYml} TEXT column.
  8. Activity row — emit AGENT_FILE_EDITED with details.diff (truncated to 5 KB) and details.runId.

8. Recipe — concrete pseudocode for AgentRunService.execute()​

async execute(context: AgentRunContext): Promise<void> {
    const agent = await this.agents.findById(context.agentId);

    // 1. Pre-flight budget check
    const remaining = await this.budgetGuard.checkBudget({
        ownerType: 'agent',
        ownerId: agent.id,
        estimatedCostCents: this.estimateCost(agent, context)
    });
    if (remaining === 'block') {
        await this.runs.markFailed(context.runId, 'Budget exceeded');
        await this.activity.record({ actionType: 'AGENT_BUDGET_EXCEEDED', userId: agent.userId, details: { agentId: agent.id } });
        return;
    }

    // 2. Load assembly inputs (parallel)
    const [files, skills, scopeContext, recentRuns, recentActivity, advancedPrompts] = await Promise.all([
        this.files.loadAll(agent),
        this.skillBindings.resolveActive({ agentId: agent.id, ... }),
        this.scope.loadDescription(agent),
        this.runs.findRecent(agent.id, 5),
        this.activity.findRecent({ agentId: agent.id, limit: 20 }),
        agent.workId ? this.advancedPrompts.findByWorkId(agent.workId) : null
    ]);

    // 3. Assemble system message (§2)
    const systemMessage = this.promptAssembler.assemble({
        files,                  // SOUL/AGENTS/HEARTBEAT/TOOLS/agent.yml
        capabilities: agent.capabilities,
        skills,                 // progressive disclosure shape
        advancedPrompts,        // when relevant
        scopeContext,
        memory: { recentRuns, recentActivity },
        kind: context.kind      // 'heartbeat' | 'task' | 'chat'
    });

    // 4. Assemble user message (§3)
    const userMessage = this.promptAssembler.userMessage(context);

    // 5. Resolve provider + model via existing facade
    const response = await this.aiFacade.createChatCompletion({
        messages: [
            { role: 'system', content: systemMessage },
            { role: 'user',   content: userMessage }
        ],
        tools: this.tools.resolveAllowedTools(agent),
        // routing:
        complexity: 'medium',   // 'simple' | 'medium' | 'complex' — Agent runs default to medium
    }, {
        userId: agent.userId,
        workId: agent.workId,
        providerOverride: agent.aiProviderId ?? undefined,
        // new field for run-level cost attribution
        agentId: agent.id,
        taskId: context.taskId
    });

    // 6. Handle tool loop (separate concern; not shown)
    await this.toolLoop.run(agent, context, response);

    // 7. Finalize
    await this.runs.markCompleted(context.runId, { summary: this.summarizer.summarize(response) });
}

9. Streaming vs non-streaming​

For heartbeats and tasks: non-streaming by default. The Agent returns a single response; no UI is waiting on incremental output. Saves a code path.

For chat replies: streaming, so the chat panel shows incremental typing. Use createStreamingChatCompletion() and pipe chunks into task_chat_messages.body updates (debounced 250 ms). The final [DONE] frame flushes the row.

10. References​