agent-loop（Agent 循环）

Agent Loop

Core rule: You must complete each phase in order. Do not skip phases. Do not report "done" until you have evidence from the Verify phase.

Phase 0 — Trigger Check

IF the task involves any of:

• reading or writing files
• running shell commands
• multi-step operations (3+ steps)
• actions that cannot be undone

THEN apply this full protocol. ELSE (pure Q&A, one-sentence reply) → skip this skill entirely.

Phase 1 — READ

Before touching anything:

• IF you need to edit a file → read it first, in this session
• IF you need to know a file's content → read it, never recall from memory
• IF you need to run a command → confirm what it does before running

FORBIDDEN: Edit a file you have not read this session
FORBIDDEN: Assume file content without reading

Phase 2 — PLAN

IF task has 3+ steps → write a numbered plan before executing:

Plan:
1. Read <file>
2. Edit <file>: change X → Y
3. Run <command> to verify
4. Report result with evidence

• IF the plan includes destructive or irreversible actions → ask for confirmation first
• Use agent-step-sequencer for plans with background processes or that must survive gateway resets
• IF agent-task-tracker is installed → save plan to memory/tasks.md

Phase 3 — EXECUTE

• Complete one step fully before starting the next
• Record progress in memory/tasks.md after each step

IF a step fails:

1. Read the full error message completely
2. Identify the root cause — do not guess
3. Change your approach before retrying

FORBIDDEN: Retry the same failing command unchanged
FORBIDDEN: Skip a failed step and continue as if it succeeded

Phase 4 — VERIFY

Before reporting done, confirm success with evidence:

FORBIDDEN: Report "done" without running a verification step
FORBIDDEN: Treat absence of error as proof of success

Phase 5 — REPORT

Report exactly three things:

1. What was done — one or two sentences
2. Verification evidence — what you checked and what it showed
3. Caveats / next steps — if any

Error Recovery Protocol

IF stuck after 2 failed attempts:

1. Stop retrying
2. Report the blocker clearly: what you tried, exact error, what you think is wrong
3. Ask the user for direction

FORBIDDEN: Silently swallow errors and report success
FORBIDDEN: Retry more than twice without changing the approach

Scope Control

Only do what was asked.

• IF you notice an unrelated bug → note it in your report, do not fix it
• IF you notice something to refactor → mention it, do not act on it

FORBIDDEN: Fix, refactor, or improve anything not mentioned in the task