Files
executus/run/ports.go
T
steve 2ef88f2a73
Adversarial Review (Gadfly) / review (pull_request) Has been cancelled
executus CI / test (pull_request) Successful in 2m21s
feat(run): InputFileStager seam — stage non-image attachments into the prompt
executus's tool.Invocation already carried InputFiles (audio/PDF/binary), but the
executor never staged them — only Images were folded into the run. This adds the
host seam mort's chat/chatbot surfaces need for audio-input parity with agentexec.

- run.Ports gains InputFiles InputFileStager (nil-safe; nil = input files silently
  ignored, run still proceeds text-only). The interface mirrors mort's skill
  FileStorage: StageInputFile(ctx, runID, agentID, name, mime, content) → file_id.
- run/input_files.go (ported from mort agentexec/input_files.go): stageInputFiles
  persists each file under run scope and appends an [ATTACHED FILES] descriptor
  block to the prompt so the agent can reach them by file_id (e.g. code_exec
  files_in → /workspace/<name>). Bytes are NEVER inlined into model context.
  Best-effort: empty/oversized(>50MB)/save-error files are skipped; colliding
  base names are disambiguated (name-2, name-3) so they don't clobber at
  /workspace/<name>.
- Executor.Run calls it after the model/toolbox build, before the loop, so the
  descriptor rides the first user turn (alongside the existing Images folding).

Tests: stages + builds the block; nil stager / no files leave the prompt intact;
dedup; empty/save-error skipping. Full suite green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-28 13:02:55 -04:00

211 lines
9.2 KiB
Go

package run
import (
"context"
"errors"
"time"
"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
"gitea.stevedudenhoeffer.com/steve/executus/deliver"
)
// ErrCriticKill is the cancellation cause the executor stamps on a run the
// critic kills, so a critic kill surfaces as a distinct "killed" status (vs a
// backstop "timeout" or a caller "cancelled"). A host CriticHandle signals a
// kill via KillCause(); the executor wraps that reason with this sentinel.
var ErrCriticKill = errors.New("run: critic killed the run")
// Ports are the host seams the run executor consumes. Every field is nil-safe:
// a light host passes the zero Ports and gets a bounded, in-memory run with no
// persistence, audit, budget, critic, delegation, or delivery — which is
// exactly a gadfly swarm task. A heavy host (mort) wires each one to a battery.
//
// This struct IS the inversion: in mort, agentexec imports agents /
// agentcritic / skillaudit and skillexec imports skills / paste directly; here
// the kernel depends only on these interfaces, and the batteries implement
// them. The mort_*_adapters.go wall becomes the set of impls.
type Ports struct {
// Audit records the run trace (start, per-step/per-tool events, final
// stats). nil = no audit.
Audit Audit
// Budget gates and meters per-caller resource use. nil = unbounded.
Budget Budget
// Critic optionally monitors a long run for hangs/runaways. nil = none.
Critic Critic
// Checkpointer persists resumable progress for durable recovery. nil = no
// checkpointing (a run interrupted by shutdown is simply lost).
Checkpointer Checkpointer
// Palette resolves SkillPalette / SubAgentPalette entries into delegation
// tools (skill__<name> / agent__<name>). nil = those entries are inert.
Palette PaletteSource
// Delivery is where the run's output + artifacts go. nil = the caller
// reads the Result in-process (the light-host default).
Delivery deliver.Delivery
// InputFiles persists non-image input attachments (audio, PDF, binary)
// carried on Invocation.InputFiles into a host file store under run scope,
// returning file_ids the agent can hand to a worker tool. nil = input files
// are silently ignored (the run still proceeds, text-only). The bytes are
// never inlined into the model context — the LLM can't read raw audio/binary.
InputFiles InputFileStager
}
// InputFileStager persists a single non-image input attachment into a host file
// store under run scope and returns a file_id the run can reference. It is the
// seam mort's skill FileStorage (and any host blob store) implements so the
// kernel can stage Invocation.InputFiles without importing a storage layer.
type InputFileStager interface {
StageInputFile(ctx context.Context, runID, agentID, name, mime string, content []byte) (fileID string, err error)
}
// RunInfo describes a run at start time — the attribution a recorder/critic
// needs. Host-neutral rename of mort's SkillRun start fields.
type RunInfo struct {
RunID string
SubjectID string // the agent/skill id being run (audit "skill_id")
Name string
CallerID string
ChannelID string
ParentRunID string
Inputs map[string]any
StartedAt time.Time
// MaxIterations is the run's base tool-dispatch step ceiling, so a critic can
// raise it relative to the baseline (see CriticHandle.MaxSteps).
MaxIterations int
}
// RunStats is the terminal roll-up a recorder's Close writes. Mirrors mort's
// skillaudit/skillexec RunStats.
type RunStats struct {
Status string // ok | error | timeout | budget_exceeded | cancelled | dry_run
Output string
Error string
ToolCalls int
RuntimeSeconds float64
InputTokens int64
OutputTokens int64
ThinkingTokens int64
}
// --- Audit ---
// Audit begins recording a run. StartRun returns a per-run RunRecorder (or nil
// to skip recording this run). The audit battery wires its Storage behind this.
type Audit interface {
StartRun(ctx context.Context, info RunInfo) RunRecorder
}
// RunRecorder records the events of one in-flight run and its final stats. It
// satisfies RunTally so the kernel can surface live token/tool counts to the
// self-status tool. Mirrors mort's skillaudit.Writer.
type RunRecorder interface {
RunTally
// OnStep records one completed agent-loop iteration's model response.
OnStep(iter int, resp *llm.Response)
// OnTool records one executed tool call + its result.
OnTool(call llm.ToolCall, result string)
// LogEvent / LogError append structured events to the run log.
LogEvent(eventType string, payload map[string]any)
LogError(msg string)
// Close writes the terminal roll-up. Detaches from the caller's context
// internally so a cancelled run still records.
Close(ctx context.Context, stats RunStats)
}
// --- Budget ---
// Budget gates and meters per-caller resource use. Mirrors mort's
// skillexec.BudgetTracker.
type Budget interface {
// Check reports whether the caller has remaining budget (nil = allowed).
Check(ctx context.Context, callerID string) error
// Commit records that the caller spent runtimeSeconds on this run.
Commit(ctx context.Context, callerID string, runtimeSeconds float64)
}
// --- Critic ---
// Critic optionally monitors a long-running run (the two-tier soft/hard
// timeout). Monitor returns a handle the executor feeds progress into and
// queries for steer/deadline decisions; a nil handle means "not monitored".
//
// The exact wiring (how the handle's Steer/Deadline bind into majordomo's
// agent.WithSteer / agent.WithMaxStepsFunc / run-context cancellation) is
// finalized in the executor; this is the seam the agentcritic battery adapts.
type Critic interface {
Monitor(ctx context.Context, info RunInfo, softTimeout time.Duration) CriticHandle
}
// CriticHandle is the executor's live link to a run's critic.
//
// Concurrency: the executor calls RecordStep/RecordToolStart/Steer from the run
// goroutine while a separate watch goroutine polls Deadline() and the run's end
// calls Stop() — so implementations MUST be safe for concurrent use across these
// methods (the critic battery's handle guards its state with a mutex).
type CriticHandle interface {
// RecordStep / RecordToolStart keep the critic's activity clock fresh so a
// healthy-but-slow run is not mistaken for a hang. RecordStep also carries the
// completed step's model response (nil-safe) so the critic's Trace can show
// what the agent actually produced, not just an iteration count.
RecordStep(iter int, resp *llm.Response)
RecordToolStart(name, args string)
// Steer returns any messages the critic wants injected into the loop (a
// nudge), drained before each step — matches majordomo agent.WithSteer.
Steer() []llm.Message
// Deadline returns the current hard-kill deadline (the critic may extend
// it); the executor binds the run context to it. Zero = no hard deadline.
Deadline() time.Time
// MaxSteps returns the current tool-dispatch step ceiling, polled by the
// executor each step (via majordomo WithMaxStepsFunc) so a critic can raise a
// healthy-but-long run's iteration budget mid-flight. Return <= 0 to defer to
// the run's base MaxIterations.
MaxSteps() int
// KillCause returns a non-nil reason iff the critic has decided to KILL this
// run (as opposed to letting the hard-deadline backstop expire). The executor
// reads it when the deadline passes: non-nil → cancel the run with
// ErrCriticKill (status "killed"); nil → the backstop expired naturally
// (status "timeout"). Hosts that never distinguish the two may return nil.
KillCause() error
// Stop ends monitoring when the run finishes.
Stop()
}
// --- Checkpointer ---
// Checkpointer persists a run's resumable progress for durable recovery.
// Mirrors mort's agentexec.RunCheckpointer.
type Checkpointer interface {
// Save persists the run's current resumable progress (throttled).
Save(ctx context.Context, st RunCheckpointState) error
// Complete clears the checkpoint on success.
Complete(ctx context.Context) error
// Fail clears the checkpoint on terminal failure. A run interrupted by
// shutdown is left untouched so boot recovery picks it up.
Fail(ctx context.Context, err error) error
}
// RunCheckpointState is the resumable snapshot a Checkpointer persists. Kept
// minimal here; the executor extends what it records during the merge.
type RunCheckpointState struct {
Messages []llm.Message
Iteration int
}
// --- PaletteSource ---
// PaletteSource resolves a RunnableAgent's SkillPalette / SubAgentPalette names
// into delegation tools and invokes them. Mirrors mort's
// SkillInvokerForPalette + AgentInvokerForPalette. nil Palette => palette
// entries are inert ("not configured" at first call).
type PaletteSource interface {
ResolveSkill(ctx context.Context, callerID, name string) (skillID string, err error)
InvokeSkill(ctx context.Context, callerID, channelID, name string,
inputs map[string]any, parentRunID string) (output, runID, status string, err error)
ResolveAgent(ctx context.Context, callerID, name string) (agentID string, err error)
InvokeAgent(ctx context.Context, callerID, channelID, name string,
prompt, parentRunID, modelTierOverride, promptPrepend string,
toolsSubset []string,
onEvent func(ctx context.Context, event, emoji string)) (output, runID, status string, err error)
}