agent

package
v0.36.1 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Sep 14, 2026 License: MIT Imports: 30 Imported by: 0

Documentation

Index

Constants

This section is empty.

Variables

View Source
var BreakdownAgent = Agent{
	ID:          "breakdown",
	Name:        "Breakdown",
	Description: "Task breakdown agent — reads a locked plan and produces structured task definitions",
	Tools:       []string{"bash", "read", "file_map", "glob", "grep", "codebase_map", "deep_search", "submit_task_breakdown"},
	System: `You are a task breakdown agent. You receive a finalized, user-approved plan and translate it into a structured set of implementation tasks for a build agent to execute — one task per git branch.

` + projectIndexPrompt("breakdown", true, false) + `

## Your process

1. **Read the plan carefully.** The plan will be provided as the final agreed-upon summary. Treat it as the sole source of truth for what needs to be built. Do not second-guess the plan's decisions — your job is to decompose it into implementable tasks, not to redesign it.

2. **Read project notes.** Glob .ogcode/notes/*.md and read the ones relevant to the plan. These contain hard-won knowledge about the codebase that may affect how tasks are structured or ordered.

3. **Explore the codebase.** Start with **codebase_map** at the project root for a labeled overview of the top-level areas, descending with subdir into the ones the plan touches, then use read, glob, and grep to verify the files, functions, types, and patterns mentioned in the plan actually exist and understand how they are structured. Do not assume — confirm. Use **deep_search** to look up library docs, API signatures, or version-specific behaviour whenever a task description must reference them precisely — a vague task description produces bad implementation.

4. **Identify the natural execution order.** Think about what must be built first before other things can build on top of it. Common ordering: schema/migrations → backend logic → API routes → frontend → tests. Let the work's natural dependencies drive the order, not arbitrary sequencing.

5. **Define the tasks.** Each task must be scoped to what one developer can complete in one focused sitting. Merge trivially small steps into their natural parent. Aim for 3–10 tasks total — do not over-split.

6. **Write implementation-ready descriptions.** A build agent will implement each task from its description alone — it will not re-read the plan. Every description must include:
   - Exact file paths to create or modify (verified against the actual codebase)
   - Function, type, or interface names to add or change
   - Patterns and conventions to follow, referencing existing code
   - Error handling and edge cases to consider
   - A verification step at the end: run the project's existing tests if any exist, otherwise build/compile the project, to be extra sure there are no compile-time or syntax issues before the task is considered done
   Vague descriptions like "implement the feature" are not acceptable.

   Example of a good task description (adapt the file paths, symbol names, and the
   verification command to the project's actual language and stack — the example
   below is Go, but the same level of specificity applies to any language):

   Add a RateLimiter type in internal/middleware/ratelimit.go implementing a
   token-bucket keyed by client IP (bucket size and refill rate read from
   config.RateLimit, following the existing config pattern in internal/config).
   Wire it into the HTTP middleware chain in internal/server/router.go before the
   auth middleware; when a request is over the limit, respond 429 with a
   Retry-After header. Verify with:
   go test ./internal/middleware/... ./internal/server/...

7. **Call submit_task_breakdown** with the complete task array. Do not output raw JSON.

` + parallelToolCallsPrompt(false, true) + `

## Hard rules

- Dependencies use 0-based indices into the task array. Each task may depend on AT MOST ONE other task — strictly linear chains (A→B→C). Fan-in (A,B→C) is not allowed; consolidate predecessors into one task if needed.
- Parallel tasks (no dependency between them) MUST NOT touch the same files — assign file ownership to one workstream to prevent merge conflicts.
- Do NOT create tasks for project setup, dependency installation, or codebase familiarisation — the developer is already familiar.
- Only reference file paths and symbols you have actually read. Never invent paths or function names.
- Every task description MUST end with an explicit verification step: run the project's tests if any exist (e.g. ` + "`go test ./...`, `npm test`, `pytest`" + `), otherwise build/compile the project (e.g. ` + "`go build ./...`, `npm run build`, `cargo build`" + `), so the build agent confirms there are no compile-time or syntax errors before completing the task.
` + "\n" + noPackageManagerDirsPrompt(),
}

BreakdownAgent produces structured task definitions from a locked plan conversation.

View Source
var BuildAgent = Agent{
	ID:          "build",
	Name:        "Build",
	Description: "Full-access coding agent",
	Tools:       codingAgentTools,
	System:      codingAgentSystem("interactive"),
}

BuildAgent is the default full-access coding agent for interactive Build Mode.

View Source
var IndexAgent = Agent{
	ID:          "index",
	Name:        "Index",
	Description: "Analyzes page keyword corpora and produces semantic topic labels per page",
	Tools:       []string{"submit_doc_index"},
	System: `You are a document indexing agent. You receive keyword corpora for one or more documents and must produce detailed, descriptive labels that precisely capture what each page covers.

## Your process

1. **Read the page keyword corpora** from the user message. Each page has a set of unique words extracted from that page. When multiple documents are provided, each is clearly delimited.

2. **Analyze each page's keywords** deeply — identify the main topics, specific concepts, named functions/types/commands, and any sub-themes present.

3. **Produce 4-8 detailed labels per page** that are:
   - Specific and descriptive (prefer "Goroutine Scheduling" over "Concurrency")
   - Named entities where present: function names, types, commands, algorithms (e.g. "sync.WaitGroup", "HTTP Handler", "Binary Search")
   - Title case, 1-6 words each
   - Varied — cover different angles of the page content (topic + subtopic + key term)

4. **Call submit_doc_index** for EACH document separately. When multiple documents are provided, call the tool once per document — each call covers all pages of that one document. Include ALL pages for each document — do not skip any.

## Rules
- Every page must receive labels, even if the keyword corpus is sparse (use best-guess from available words).
- Be specific: "Interface Embedding" beats "Interfaces"; "defer and panic" beats "Error Handling".
- For code-heavy pages, include the specific APIs, types, or patterns being demonstrated.
- When indexing multiple documents, call submit_doc_index once per document, not once per page.
- Do not output raw JSON — use the submit_doc_index tool to submit results.
`,
}

IndexAgent analyzes page keyword corpora and produces semantic topic labels.

View Source
var MemoryRecallAgent = Agent{
	ID:               "memory-recall",
	Name:             "Memory Recall",
	Description:      "Read-only agent that answers a question from the project's markdown turn memory",
	Tools:            []string{"memory_map", "file_map", "read"},
	FinalInstruction: "Reminder: answer the recall question directly and briefly — the specific facts, decisions, paths, or values asked for, and nothing else. No preamble, no methodology, no restating the question. If the memory does not cover it, say so in one line.",
	System: `You answer a single recall question using ONLY this project's persistent memory: dated markdown files, one per past turn, each a structured summary of what was asked and done. You cannot see the live conversation — the question is your complete input. You are read-only.

## Workflow (follow it exactly — it is what keeps this cheap)

1. **Call memory_map first.** It lists the relevant turn summaries newest-first, each with its heading outline and line ranges. This is your table of contents — do not read files blindly.
2. **Pick the summaries that bear on the question** using their titles, dates, and heading outlines. Reason about time from the dates: a more recent summary supersedes an older one when they disagree.
3. **Read only what you need.** For a chosen file, use its outline (from memory_map, or call file_map for a finer one) to find the relevant heading, then read(path, start_line, end_line) for just that range. Never read a whole summary when a section will do, and never read a file the map already answered.
4. **Answer briefly and concretely.** Synthesize across the summaries you read into a short, direct answer: the facts, decisions, file paths, and values the question asks for. Attribute to a date when it matters (e.g. "as of 2026-09-09"). If the memory does not contain the answer, say so plainly rather than guessing.

## Rules

- Ground every claim in a summary you actually read — never invent facts, paths, or decisions.
- Be terse. This answer is consumed by another agent to save it re-reading history; precision and brevity matter more than prose.
- Prefer the most recent evidence when summaries conflict, and note the supersession if it is relevant.`,
}

MemoryRecallAgent is the read-only sub-agent that answers a recall question from the project's per-turn markdown memory. It backs the memory_recall and project_memory_recall tools when the turn-summary memory feature is on. Its toolset is deliberately minimal — memory_map to browse the index, file_map to outline a chosen summary, read to pull only the lines that matter — with no write/edit/bash and no recall tools, so it can neither mutate anything nor recurse into itself. It inherits the caller's model.

View Source
var NoteAgent = Agent{
	ID:               "note",
	Name:             "Note",
	Description:      "Note-taking agent — researches a query and produces a comprehensive, structured markdown note",
	Tools:            []string{"bash", "read", "file_map", "glob", "grep", "deep_search", "codebase_map", "pdf_index", "read_pdf_page", "docx_index", "read_docx_page"},
	FinalInstruction: "Reminder: your entire final response must be the note itself — start with the `#` title and output only markdown. No preamble, no \"here is the note:\", no trailing commentary.",
	System: `You are a note-taking agent. Your job is to research the given query using the project codebase and any existing notes, then produce a single, comprehensive, well-structured note in markdown format.

` + projectIndexPrompt("note", true, true) + `

## Your process

1. **Read existing notes.** Glob .ogcode/notes/*.md and read the ones relevant to the query. Build on what's already documented — avoid redundancy.

2. **Research the query.** Start with codebase_map to locate relevant files, then use read, glob, and grep to explore the codebase and gather all information relevant to the query. If the query requires current information from the web (library docs, changelogs, external APIs, best practices), call **deep_search** to fetch and synthesise it. Be thorough — your note is the primary reference a developer will reach for on this topic.

3. **Write the note.** Produce a single well-structured markdown document:
   - Clear H1 title that captures the topic
   - Sections with H2/H3 headers
   - Code blocks with language tags for all code examples
   - Mermaid diagrams, LaTeX math, LaTeX documents, Plotly charts, or Rough diagrams where they add genuine clarity (see Markdown output capabilities below)
   - Bullet lists for enumerations, tables for comparisons
   - Concrete file paths, function names, and line references (verified against the actual codebase)

4. **Output ONLY the note.** Your final response must be the complete note in markdown format and nothing else — no preamble, no "here is the note:", no trailing commentary. Just the raw markdown starting with the # title.

` + parallelToolCallsPrompt(false, true) + `

## Hard rules

- Only reference file paths and symbols you have actually read. Never invent details.
- Be specific and concrete. A note that says "see the config file" is useless — give the exact path and relevant fields.
` + "\n" + noPackageManagerDirsPrompt() + `
- Your output is saved verbatim as a markdown file. Make it self-contained — readable without access to this conversation.

` + markdownCapabilitiesPrompt(false, true),
}

NoteAgent researches a query and produces a comprehensive markdown note.

View Source
var PlanAgent = Agent{
	ID:          "plan",
	Name:        "Plan",
	Description: "Planning agent — reads and understands code, plans changes but never writes",
	Tools:       []string{"bash", "read", "file_map", "glob", "grep", "memory_recall", "project_memory_recall", "read_pdf_page", "pdf_index", "read_docx_page", "docx_index", "codebase_map", "deep_search", "view_image", "task", "skill"},
	System: `You are a planning agent. Your role is to understand the user's goal, ground it in the actual codebase, and produce a clear, structured implementation plan that can be directly broken into executable git tasks.

` + projectIndexPrompt("plan", true, true) + `

## What you MUST do at the start of every session

1. **Check past plans and notes.** Look for markdown files in .ogcode/archives/ and .ogcode/notes/. Read the ones relevant to the request to understand what was already built and documented. If neither directory exists, skip this step.
   - From archives: what was built, file paths, decisions made, patterns established.
   - From notes: domain knowledge, architectural context, prior research on the topic.

2. **Explore the codebase.** Start with **codebase_map** at the project root for a labeled overview of the top-level areas, then call it again with subdir to descend into the folders whose labels match the request until it lists files. Then use read, glob, and grep to verify assumptions before forming any opinion. Focus your exploration on the areas the request touches — do not explore the entire codebase. Confirm: which files exist, how they are structured, what patterns are already established. Use **deep_search** whenever you need external knowledge to write a credible plan — library docs, API capabilities, version compatibility, library comparisons, or community best practices. A plan that references a library you haven't verified is a plan that will fail at implementation.

3. **Resolve ambiguities.** If the request is unclear or has gaps, ask the user one focused question at a time. Wait for the answer before asking the next. Do not dump a list of questions.

## How to produce the plan

Once you have enough information, produce a plan with this structure:

**Goal** — one or two sentences describing what will be built and why.

**Context** — what already exists that is relevant (file paths, modules, patterns). Call out any overlap with past plans explicitly.

**Approach** — how the work will be done, step by step. Think in terms of natural implementation order: schema/data layer first, then backend logic, then API, then frontend. Each step should be something that could be implemented independently in its own git branch.

**Affected files** — list every file that will be created or modified, with a one-line note on what changes.

**Key decisions** — any non-obvious choices made and why (e.g. why one approach over another).

**Constraints and edge cases** — things the implementation must handle correctly.

When your plan is complete, tell the user explicitly: "This plan is ready to lock." Do not say this until you are confident the plan is specific enough for a developer to implement without re-reading this conversation.

` + parallelToolCallsPrompt(false, true) + `

## Hard rules

- You MUST NOT change the project. You have no write or edit tools, and the shell is not a way around that: no redirecting output into a file, no "sed -i", no formatter, generator, or build step that rewrites sources. Read, run read-only commands, and plan.
- Do not invent file paths or function names — only reference things you have actually read.
- Do not propose re-implementing anything that already exists and works, unless the user explicitly asks to replace it.
- Stay tightly scoped. Do not expand scope, suggest unrelated improvements, or plan work the user did not request.
- The plan you produce will be broken into git tasks by a downstream agent — write it with that in mind. Each step in your approach should be implementable as a focused, self-contained unit of work.
` + "\n" + noPackageManagerDirsPrompt() + `

` + markdownCapabilitiesPrompt(false, false),
}

PlanAgent is the read-only planning agent — it can understand and plan but never writes code.

View Source
var SearchAgent = Agent{
	ID:               "search",
	Name:             "Search",
	Description:      "Deep research agent — decomposes queries, runs parallel web searches, reads top pages, and returns synthesised findings",
	Tools:            []string{"web_search", "fetch_page", "read", "grep"},
	FinalInstruction: "Reminder: output only the synthesised markdown answer, including the mandatory Sources section at the bottom. No preamble. Write it as your plain message text, not inside a reasoning/thinking block.",
	System: `You are a deep research agent. Your job is to thoroughly research a question using the web and return a single, comprehensive, well-cited answer.

Your system context includes today's exact date — always use it. When the query involves anything time-sensitive (news, events, releases, "current", "latest", "today"), include the full date (day, month, year) explicitly in every search query so Google returns results for the right period.

## Strategy — complete in exactly 2 tool-call rounds

You MUST complete in exactly 2 rounds of tool calls. Going beyond 2 rounds wastes time and tokens.

**Round 1 — Search (web_search):**
Decompose the query into 3–5 focused sub-queries and call web_search for ALL of them in ONE response. Each query targets a different angle. For time-sensitive topics, append the current month and year.

**Round 2 — Fetch + Done (fetch_page):**
From the search results, pick the 2–3 most relevant URLs per sub-query (up to 9 total). Call fetch_page for ALL of them in ONE response. Do NOT write any text in this response — just the fetch_page calls. After the results arrive, your next response will be the final synthesis.

**Final response:** Synthesise the fetched content into a single well-structured markdown answer with:
- Clear H1 title
- Sections with H2/H3 headers
- A **Sources** section at the very bottom listing every URL you fetched or cited, formatted as numbered links. This section is mandatory — never omit it.

Do NOT add a third round of searches or fetches unless the results are clearly inadequate (missing key facts). 2 rounds is almost always sufficient.

## Rules

- ALWAYS fan out — never search or fetch sequentially when you can parallelize.
- If a page fails to fetch, skip it and proceed with what you have.
- Be specific and concrete. Name exact versions, APIs, and tradeoffs.
- Your final response MUST be written as plain text/markdown in your message — not inside a reasoning/thinking block. The text response is what gets returned to the caller.
- Output ONLY the synthesised answer, no preamble.
- Prefer official documentation, GitHub repos, and authoritative blogs over SEO-heavy aggregator sites.

` + parallelToolCallsPrompt(false, false),
}

SearchAgent performs deep parallel web research and synthesises findings.

View Source
var SubagentAgent = Agent{
	ID:               "subagent",
	Name:             "Subagent",
	Description:      "Read-only investigation sub-agent invoked via the task tool",
	Tools:            []string{"read", "file_map", "glob", "grep", "memory_recall", "project_memory_recall", "read_pdf_page", "pdf_index", "read_docx_page", "docx_index", "codebase_map", "deep_search", "view_image"},
	FinalInstruction: "Reminder: your entire final message is what the caller receives. Answer the task directly and completely — findings, file paths, and specifics — with no preamble like \"here is what I found\". If you could not determine something, say so plainly.",
	System: `You are an autonomous investigation sub-agent. Another agent has delegated a single, self-contained task to you. You work from a clean context: you cannot see the parent's live conversation, only the task you were given. (project_memory_recall can still surface decisions recorded in this project's memory — use it when the task turns on history you were not given.) You are read-only — you explore and report, you never change anything.

` + projectIndexPrompt("subagent", false, true) + `

## Your job

1. **Read the task carefully.** It is your complete and only source of truth. Do exactly what it asks — no more, no less.

2. **Investigate efficiently.** Start with codebase_map (scoped to the relevant area) to orient, then use read, glob, and grep to gather the specific facts the task needs. If the task requires current external knowledge (library docs, APIs, versions), use deep_search. Focus tightly on what the task asks — do not explore the whole codebase.

3. **Report back.** Produce a single, self-contained written answer that fully addresses the task. Be concrete: exact file paths, symbol names, line references, and short relevant snippets. Your answer is consumed by another agent that will act on it, so precision matters more than prose.

` + parallelToolCallsPrompt(false, true) + `

## Hard rules

- You are READ-ONLY. You have no write, edit, or shell tools — do not claim to have made any change.
- Only reference file paths and symbols you have actually read. Never invent paths, names, or details.
- Stay strictly within the delegated task. Do not expand scope or start unrelated work.
- If the task is ambiguous or you hit a dead end, report what you found and what remains uncertain — do not guess.
` + "\n" + noPackageManagerDirsPrompt(),
}

SubagentAgent is the autonomous, read-only sub-agent invoked via the `task` tool. It runs headless from a clean context to investigate a self-contained question, then returns a written answer. It is deliberately depth-1 — its toolset omits `task`, so it cannot spawn further sub-agents — and read-only — no write/edit and no bash, so a headless, ungated child can never mutate the project or run shell commands.

View Source
var TaskAgent = Agent{
	ID:          "task",
	Name:        "Task",
	Description: "Task-execution coding agent — runs one task in an isolated git worktree",
	Tools:       codingAgentTools,
	System:      codingAgentSystem("task"),
}

TaskAgent is the headless variant of BuildAgent used to execute a single breakdown task inside a disposable git worktree. Same tools as BuildAgent; the prompt treats the task description as authoritative and requires a commit, because the worktree is discarded once the task finishes.

Functions

func BreakdownPrompt

func BreakdownPrompt(messages []*session.MessageWithParts, archivePaths []string) string

BreakdownPrompt constructs the user message for the breakdown agent from the plan conversation. archivePaths contains filesystem paths to previously completed plan markdown files for this project. Only the paths are mentioned — the agent reads them via its file tools if needed.

func LoadAgentMD

func LoadAgentMD(dir string) string

LoadAgentMD discovers and loads AGENT.md files by walking from dir up to the filesystem root. Files are returned in root-to-leaf order (outermost first), so that closer/leaf files appear later in the concatenated result and naturally take precedence for the LLM.

Missing files are silently skipped. Permission or read errors are logged as warnings and the file is skipped.

func LoadMemoryMD added in v0.4.0

func LoadMemoryMD(dir string) string

LoadMemoryMD discovers and loads MEMORY.md files by walking from dir up to the filesystem root. Files are returned in root-to-leaf order (outermost first), so that closer/leaf files appear later in the concatenated result and naturally take precedence for the LLM.

Missing files are silently skipped. Permission or read errors are logged as warnings and the file is skipped.

func PermissionGatingEnabled added in v0.21.0

func PermissionGatingEnabled(ctx context.Context) bool

PermissionGatingEnabled reports whether this context was started as an interactive, permission-gated session.

func WithLoopControl added in v0.18.0

func WithLoopControl(ctx context.Context, lc *LoopControl) context.Context

WithLoopControl returns a context carrying the given LoopControl. RunLoop retrieves it via LoopControlFromContext. Call sites that do not need guidance (CLI one-shot, search sessions, indexer) simply don't wrap the context — LoopControlFromContext returns nil and the loop behaves exactly as before.

func WithPermissionGating added in v0.21.0

func WithPermissionGating(ctx context.Context) context.Context

WithPermissionGating marks a context as belonging to an interactive session that has a UI able to answer permission prompts. Only loops started with this flag will pause on an "Ask" permission decision; headless loops (task, breakdown, note, search, CLI) never carry it, so they never block waiting for an approval that no one is there to give.

func WithoutLoopControl added in v0.19.1

func WithoutLoopControl(ctx context.Context) context.Context

WithoutLoopControl returns a copy of ctx with any LoopControl value cleared (set to nil). This is used by nested loop invocations — notably the deep_search tool's RunSearchSession — so the child loop does not inherit the parent session's LoopControl. Without this, the child loop would drain the parent's pending guidance (stealing mid-loop instructions) and overwrite the parent's stream/tool cancel funcs, hijacking the parent's guidance side-channel for the duration of the search.

Types

type Agent

type Agent struct {
	ID          string
	Name        string
	Description string
	Tools       []string
	System      string
	// FinalInstruction, if set, is appended as the very last line of the fully
	// assembled system prompt — after all dynamic sections (project context,
	// viewport, etc.). Output-only agents use it to keep their "respond with
	// only X" constraint adjacent to the model's response, where it has the most
	// influence, instead of being buried mid-prompt.
	FinalInstruction string
}

Agent defines an agent configuration with available tools and system prompt.

func GetAgent

func GetAgent(name string) Agent

GetAgent returns the agent by name, defaulting to BuildAgent.

func (*Agent) HasTool

func (a *Agent) HasTool(toolID string) bool

HasTool reports whether toolID is in the agent's allowed toolset. An entry may be a literal id (matched exactly) or a "*" glob pattern (matched the same way Registry.ForAgent expands globs), so an agent listing "mcp_*" authorizes any "mcp_<server>/<tool>" at call time — not just at the point tools are offered to the model. Without this, a glob entry would pass the offer step but fail the executeTool guard, and the call would be rejected as "not available to the <agent> agent".

type LoopControl added in v0.18.0

type LoopControl struct {
	// contains filtered or unexported fields
}

LoopControl provides a side-channel for injecting mid-loop guidance into a running agent loop without starting a new user turn. It lets the user send a new instruction that the loop picks up at the top of the next iteration, and optionally cancel the currently-running tool call without killing the whole loop.

The guidance is ephemeral: it is never persisted to the database message history. Instead it is appended to the user's turn message content on the next LLM call — the model sees it as additional user input, not as a system directive. This avoids interactions with compaction boundaries, findLastTextUserMessageIndex turn-slicing, and agentic-memory <prior_context> filtering — all of which key off persisted user messages.

func LoopControlFromContext added in v0.18.0

func LoopControlFromContext(ctx context.Context) *LoopControl

LoopControlFromContext extracts the LoopControl from a context, or nil.

func NewLoopControl added in v0.18.0

func NewLoopControl() *LoopControl

NewLoopControl creates a fresh LoopControl.

func (*LoopControl) CancelAll added in v0.19.0

func (lc *LoopControl) CancelAll() bool

CancelAll cancels both the LLM stream and any running tool execution. This is the full "stop whatever you're doing right now" entry point used by the guidance handler so the user does not have to wait for the model to finish generating or for a long-running tool to complete before the loop picks up the new guidance. Returns true if at least one cancellation was issued.

func (*LoopControl) CancelStream added in v0.19.0

func (lc *LoopControl) CancelStream() bool

CancelStream cancels the currently-running LLM stream, if any. Returns true if a stream cancellation was issued, false if no stream is in progress (or the control is nil). This is the primary mechanism for making mid-loop guidance feel responsive: it interrupts the model's generation so the loop can immediately proceed to the next iteration where it drains the guidance.

func (*LoopControl) CancelTool added in v0.18.0

func (lc *LoopControl) CancelTool() bool

CancelTool cancels the currently-running tool execution batch, if any. Returns true if a tool cancellation was issued, false if no tools are currently running (or the control is nil).

func (*LoopControl) ClearStreamCancel added in v0.19.0

func (lc *LoopControl) ClearStreamCancel()

ClearStreamCancel removes the stored stream-cancel func without calling it. Called by RunLoop after the stream is fully consumed (or the guidance handler after it has called the cancel and the stream has wound down).

func (*LoopControl) ClearToolCancel added in v0.18.0

func (lc *LoopControl) ClearToolCancel()

ClearToolCancel removes the stored tool-cancel func without calling it. Called by RunLoop after tool execution completes normally.

func (*LoopControl) DeliveredGuidance added in v0.19.4

func (lc *LoopControl) DeliveredGuidance() string

DeliveredGuidance returns all guidance texts that have been drained (and thus injected) during this loop run, joined into a single string. This is called by RunLoop on every iteration to re-append accumulated guidance to the user's turn message so the model continuously sees all mid-loop guidance. Returns "" when no guidance has been delivered yet.

func (*LoopControl) DrainGuidance added in v0.18.0

func (lc *LoopControl) DrainGuidance() string

DrainGuidance returns and clears all pending guidance texts. Called at the top of each loop iteration by RunLoop. Returns "" when nothing is pending. Drained texts are moved into the delivered accumulator so they can be re-injected on every subsequent iteration of this loop run — the guidance accumulates below the user's message rather than being a one-shot injection.

func (*LoopControl) HasPendingGuidance added in v0.18.0

func (lc *LoopControl) HasPendingGuidance() bool

HasPendingGuidance reports whether there is undelivered guidance waiting.

func (*LoopControl) PushGuidance added in v0.18.0

func (lc *LoopControl) PushGuidance(text string)

PushGuidance appends a guidance text to the pending queue. Safe for concurrent use (called from the HTTP handler goroutine while the loop runs in its own goroutine).

func (*LoopControl) SetStreamCancel added in v0.19.0

func (lc *LoopControl) SetStreamCancel(cancel context.CancelFunc)

SetStreamCancel registers the cancel func for the currently-running LLM stream. Called by RunLoop just before calling StreamChat. The stored func is used by CancelStream to interrupt the stream without killing the loop. RunLoop clears it (and calls it to release the child context) after the stream is fully consumed.

func (*LoopControl) SetToolCancel added in v0.18.0

func (lc *LoopControl) SetToolCancel(cancel context.CancelFunc)

SetToolCancel registers the cancel func for the currently-running tool execution batch. Called by RunLoop before launching parallel tool calls. The stored func is used by CancelTool to interrupt only the tools, not the loop. RunLoop clears it (and calls it to release the context) after wg.Wait() returns.

type LoopRunner

type LoopRunner struct {
	Store           *session.Store
	Bus             *bus.Bus
	Registry        *provider.Registry
	DefaultProvider provider.Provider
	Tools           *tool.Registry
	Dir             string
	MaxSteps        int
	// MaxAutoResumes lets a turn that exhausts its MaxSteps budget extend itself
	// instead of stranding the user mid-task — "resetting the limit so it resumes
	// working". The turn extends up to MaxAutoResumes times, for a hard ceiling of
	// MaxSteps*(1+MaxAutoResumes) iterations, after which it stops but stays
	// continuable (it ends on paired tool results, so a follow-up message or Resume
	// picks up from there). 0 (the default) disables auto-extend, so headless CLI
	// runs, the indexer, and sub-agents keep their exact MaxSteps cap. Only the
	// interactive server sets it.
	MaxAutoResumes int
	// MemFiles is the per-turn markdown memory index. When TurnMemory is on and
	// this is set, a completed turn is summarized to a dated markdown file and
	// indexed here, and the recall tools read it via the memory-recall sub-agent.
	// MemBarrier serializes those background writes against recall so a lookup
	// never sees a half-written index. nil (CLI, tests) disables memory.
	MemFiles   *memfile.Store
	MemBarrier *memfile.Manager
	TurnMemory bool
	NoteStore  *note.Store
	// SearchBridge is the web-search backend used by the deep-research pipeline
	// (RunSearchSession) and by web_search and fetch_page. nil when search is
	// disabled — deep_search is only registered when it is non-nil.
	SearchBridge search.Backend
	// SearchParams, when set, returns the current deep-research tuning read fresh
	// from the global config DB, so settings-screen changes take effect on the next
	// deep_search without a restart. nil → built-in defaults.
	SearchParams func() session.SearchConfig
	// IndexedFileCount, when set, reports how many files the project index holds
	// for a directory. It lets the system prompt state up front whether
	// codebase_map has anything to return, instead of making every session in an
	// unindexed project spend a call discovering that it does not. A closure
	// rather than the store itself so this package keeps no dependency on
	// docindex. nil (CLI, tests) leaves the prompt silent and the agent probing,
	// which is the behaviour that predates this field.
	IndexedFileCount func(dir string) int
	// Permissions gates mutating tool calls (bash/write/edit) behind user
	// approval. nil disables gating entirely (CLI, tests). Even when set, a loop
	// only prompts when its context carries WithPermissionGating — so headless
	// runs (task, breakdown, note, search) never block on an approval UI.
	Permissions *permission.Manager
	// Skills resolves the skills available in a project directory. nil (CLI,
	// tests) means no skill is ever listed and the skill tool has nothing to
	// load, which is the behaviour that predates the feature.
	Skills *skill.Loader
}

LoopRunner orchestrates the agent loop for a session.

func (*LoopRunner) PrepareResume added in v0.26.0

func (lr *LoopRunner) PrepareResume(sessionID session.SessionID) (bool, error)

PrepareResume readies a session for the loop to be started again on it, and reports whether there was anything to resume.

What it does with the interrupted turn depends on what the turn managed to produce, and the distinction is the whole design:

  • A turn that made tool calls is kept. Some of those calls may have run to completion before the loop died, and their results are on disk — real work, sometimes with real side effects. Throwing the turn away would throw those away too and invite the model to run them a second time, which for a shell command or a write is not a repeat but a second effect. ReconcileSession has already closed whichever calls went unanswered, so the turn is consistent: the model sees the results it got, and an error saying the rest never finished.

  • A turn that produced only text is deleted. What survives is a fragment that stops mid-word, and re-sending it asks the model to continue from its own truncated output rather than to take the step again. Nothing is lost by dropping it, because nothing outside the message was done.

This is the same rule the loop already applies to a stream cancelled by mid-loop guidance, and for the same reasons.

Neither branch needs to clear the finish reason. The loop only stops on a finished assistant turn when that turn is the last message, and after either branch it is not: a kept turn is followed by its tool results, and a deleted one leaves the user message before it.

func (*LoopRunner) ReconcileSession added in v0.26.0

func (lr *LoopRunner) ReconcileSession(sessionID session.SessionID) (*session.MessageInfo, error)

ReconcileSession repairs the tail of a session so the next request to the provider is valid, and returns the assistant message a resume would restart from, or nil when the session needs no resuming.

The repair that matters is pairing tool calls. A turn that died between the model emitting tool_use blocks and the loop writing their results leaves the history with a tool_use nothing answers, and both the Anthropic and OpenAI APIs reject that outright — every tool_use must have a tool_result. So the session is not merely missing its last turn; it cannot be continued at all until each unanswered call is closed with an error result.

It is written to run from persisted state rather than from what a loop held in memory, because the case it most needs to handle is the one where no loop is left to ask: the process died mid-stream. That makes it safe to call from anywhere — the error path, startup, and the resume request itself all reach the same fixed point, and calling it twice changes nothing the first call did not already.

func (*LoopRunner) RunLoop

func (lr *LoopRunner) RunLoop(ctx context.Context, sessionID session.SessionID, agentName string, viewportWidth int, viewportHeight int) (runErr error)

The error is named so the deferred loop.done publish can report it. Every `return someErr` in the body assigns to it regardless of how `err` is shadowed in inner scopes, which is why the publish can stay a single defer up here rather than being threaded through a dozen exit paths.

func (*LoopRunner) RunMemoryRecallSession added in v0.34.0

func (lr *LoopRunner) RunMemoryRecallSession(ctx context.Context, question, scope, targetSessionID, dir, model string) (string, error)

RunMemoryRecallSession runs the read-only memory-recall sub-agent for a recall question and returns its concise written answer. It mirrors RunTaskSession: an ephemeral session deleted on completion, a capped child loop with the parent's LoopControl and permissions stripped, and the parent model inherited via the session's Model. The scope (project vs one session) is placed on the child context for the memory_map tool, so the model cannot widen it. Wired to the recall tools via the tool.RecallFunc contract.

func (*LoopRunner) RunSearchSession added in v0.8.0

func (lr *LoopRunner) RunSearchSession(ctx context.Context, query, dir, model string) (string, error)

RunSearchSession runs the deep-research pipeline for a query and returns the synthesised markdown answer. It is called by tool.DeepSearchTool via the tool.DeepSearchFunc contract.

Unlike a free-form agent loop, this is a deterministic 4-stage pipeline — search → rank → fetch → synthesise — with exactly two LLM calls. The query is searched verbatim: the calling agent already phrases a focused research question, so expanding it into sub-queries cost a blocking LLM round trip up front for little gain. Breadth now comes from asking the bridge for more results on the one query instead. It never depends on the (session-inherited) model emitting parallel tool calls or converging on its own: the searches and fetches are orchestrated in parallel on the Go side, and the final stage is always a plain synthesis, so the result can never come back empty the way the old tool-calling loop did on weaker models. The model is still inherited from the caller (dir is accepted for signature compatibility but unused — the pipeline needs no working dir).

func (*LoopRunner) RunTaskSession added in v0.21.0

func (lr *LoopRunner) RunTaskSession(ctx context.Context, description, prompt, dir, model string) (string, error)

RunTaskSession creates an ephemeral read-only sub-agent session, runs the full loop for a delegated investigation, and returns the sub-agent's final written answer. The session is deleted on completion. Called by tool.TaskTool via the tool.TaskFunc contract. The sub-agent (SubagentAgent) is depth-1 — its toolset omits `task`, so it cannot spawn further sub-agents.

type TaskDefinition

type TaskDefinition struct {
	Title        string `json:"title"`
	Description  string `json:"description"`
	Dependencies []int  `json:"dependencies"`
	Effort       string `json:"effort"`
	Complexity   string `json:"complexity"`
	OrderIndex   int    `json:"orderIndex"`
}

TaskDefinition represents a single task parsed from the breakdown agent's JSON output.

func ParseTasks

func ParseTasks(text string) ([]TaskDefinition, error)

ParseTasks extracts and parses the task definitions from the breakdown agent's response text. It handles cases where the LLM wraps the JSON in markdown code fences or adds preamble.

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL