Copy-paste messages for every operator intervention. When you need to correct an agent, find the intervention here, paste the message, and verify.
- Detect the issue (monitoring dashboard, terminal output, git status)
- Find the matching intervention below
- Copy the MESSAGE template
- Replace all
[bracketed]values with specifics - Paste into the agent's terminal
- Wait for the agent to respond
- VERIFY the intervention worked
Escalation ladder: 5 min observe → 15 min intervene → 30 min reassign. See REACTIONS.md for decision trees.
DETECT: git diff --name-only shows files outside agent's territory.
MESSAGE:
STOP. You modified [FILE_PATH] which is outside your territory.
Your territory: [TERRITORY_LIST]
The file you modified belongs to [OTHER_AGENT].
Revert now: git checkout -- [FILE_PATH]
Continue with your remaining chains within your territory.
VERIFY: git diff --name-only shows only files within the agent's territory.
SEVERITY: YELLOW
DETECT: New file created outside territory (check git status for untracked files).
MESSAGE:
HOLD. You created [FILE_PATH] which is outside your territory.
Your territory: [TERRITORY_LIST]
Delete this file: rm [FILE_PATH]
If you need this file for your chain, document it as a blocker:
"Need [FILE_PATH] created — assign to [CORRECT_AGENT]."
Continue with your chains.
VERIFY: File is removed. No untracked files outside territory. SEVERITY: YELLOW
DETECT: Agent's completion report claims changes to files it hasn't actually modified (or vice versa).
MESSAGE:
Your completion report lists [FILE] as modified, but git shows no changes to it.
OR: git shows changes to [FILE] but your report doesn't mention it.
Update your completion report to accurately reflect:
- Files actually modified (check: git diff --name-only)
- Lines changed per file
- Reason for each change
VERIFY: Completion report matches git diff --name-only exactly.
SEVERITY: GREEN
DETECT: Agent starts Chain 3 before completing Chain 1 or Chain 2.
MESSAGE:
STOP. You started Chain [N] before completing Chain [N-1].
Chain execution protocol: Complete each chain fully before starting the next.
Order: P0 → P1 → P2 → P3. Within same priority: sequential order.
Go back to Chain [N-1]. Complete it (trace → fix → verify → document).
Then proceed to Chain [N].
VERIFY: Agent resumes the correct chain and completes it before moving on. SEVERITY: YELLOW
DETECT: Agent marks a chain as complete but didn't run build/tests or verify the fix.
MESSAGE:
HOLD. You marked Chain [N] as complete but did not verify.
Before moving to Chain [N+1], you MUST:
1. Run build: [BUILD_COMMAND]
2. Run tests: [TEST_COMMAND]
3. Confirm the original signal is resolved: [VERIFICATION_STEP]
Do this now. Do not start Chain [N+1] until Chain [N] is fully verified.
VERIFY: Agent runs build + tests, reports results. SEVERITY: YELLOW
DETECT: Agent starts working on an unrelated task before finishing the current chain.
MESSAGE:
STOP. You're context-switching mid-chain.
Current chain: Chain [N] — [CHAIN_TITLE]
You started working on: [UNRELATED_WORK]
Revert the unrelated changes. Return to Chain [N].
Complete it fully (fix → verify → document) before doing anything else.
Exception: If you discovered a P0 issue, document it and flag for operator.
VERIFY: Agent reverts unrelated work and resumes the current chain. SEVERITY: YELLOW
DETECT: Agent is repeatedly editing the wrong file/function. Changes don't fix the signal.
MESSAGE:
STOP iterating. The root cause is not where you're looking.
You've been modifying [WRONG_FILE]. The actual issue is:
File: [CORRECT_FILE]
Line: [LINE_NUMBER]
Issue: [DESCRIPTION]
Revert your recent changes: git checkout -- [WRONG_FILE]
Read [CORRECT_FILE] starting at line [LINE_NUMBER].
The fix is: [SPECIFIC_GUIDANCE]
VERIFY: Agent reads the correct file and makes a targeted fix. SEVERITY: YELLOW
DETECT: Agent asks questions or makes assumptions about unknown context.
MESSAGE:
Here's the information you need:
[ANSWER TO AGENT'S QUESTION]
Specifically:
- [FACT 1]
- [FACT 2]
- [FACT 3]
Continue with Chain [N] using this information.
VERIFY: Agent proceeds with the correct information. No more guessing. SEVERITY: GREEN
DETECT: Agent can't make progress even after corrections. Chain requires work outside agent's capability or territory.
MESSAGE:
PARK Chain [N]. This chain is blocked and cannot be completed in this sprint.
Document in your completion report:
- Chain [N]: PARKED
- Reason: [BLOCKER_DESCRIPTION]
- Needed: [WHAT_WOULD_UNBLOCK_IT]
Move to Chain [N+1]. If all remaining chains are complete, write your
completion report and commit.
VERIFY: Agent parks the chain, documents the blocker, moves to next chain. SEVERITY: YELLOW
DETECT: Agent reports a P0 discovery that falls within their territory.
MESSAGE:
P0 ACKNOWLEDGED. This is now your highest priority.
PAUSE your current chain (Chain [N]).
NEW PRIORITY: Fix the P0 issue you discovered.
P0 Chain: [TITLE]
Vector: [TRACE_FROM_P0_REPORT]
Fix: [GUIDANCE]
Verify: [VERIFICATION]
After fixing, resume Chain [N] from where you left off.
VERIFY: P0 is fixed and verified. Agent resumes original chain. SEVERITY: RED
DETECT: Agent reports a P0 discovery in another agent's territory.
MESSAGE to discovering agent:
P0 ACKNOWLEDGED. You correctly flagged this.
This is in [OTHER_AGENT]'s territory. Do NOT fix it.
Document it in your completion report under P0 DISCOVERIES.
Continue with your assigned chains.
MESSAGE to responsible agent:
P0 ESCALATION from [DISCOVERING_AGENT].
NEW P1 CHAIN (execute before your remaining P2/P3 chains):
Title: [P0_TITLE]
Vector: [TRACE]
Signal: [DESCRIPTION]
Fix: [GUIDANCE]
Verify: [VERIFICATION]
VERIFY: Discovering agent continues their work. Responsible agent addresses the P0. SEVERITY: RED
DETECT: P0 is severe enough that continuing the sprint would make things worse (data corruption, security breach active).
MESSAGE (broadcast to all agents):
SPRINT PAUSED. P0 Critical.
ALL AGENTS: Stop your current work immediately.
1. Commit everything you have right now
2. Write a partial completion report with current state
3. Do NOT start new chains
Reason: [P0_DESCRIPTION]
I will resume the sprint after the P0 is resolved.
Stand by for further instructions.
VERIFY: All agents commit and stop. Sprint health → RED. SEVERITY: RED
DETECT: Two agents modified the same file (discovered during merge or pre-merge scan).
MESSAGE to later-merge-order agent:
Your changes to [FILE] conflict with [EARLIER_AGENT]'s changes.
[EARLIER_AGENT] has merge priority (earlier in merge order).
Their changes will be kept. You need to redo your change to [FILE]
so it works with [EARLIER_AGENT]'s version.
After [EARLIER_AGENT]'s branch is merged, rebase your branch:
git rebase main
Then re-apply your fix to [FILE] in a way that's compatible.
VERIFY: Later agent's branch rebases cleanly and the file works with both changes. SEVERITY: YELLOW
DETECT: Agent B needs a change that Agent A is still working on.
MESSAGE to blocked agent:
Chain [N] is blocked waiting on [OTHER_AGENT]'s Chain [M].
PARK Chain [N]. Move to your next chain.
I'll notify you when [OTHER_AGENT]'s work is available.
Document: "Chain [N]: PARKED — waiting on [OTHER_AGENT] Chain [M]."
MESSAGE to blocking agent (after they finish):
[BLOCKED_AGENT] is waiting on your Chain [M] output.
Please ensure Chain [M] is committed before moving on.
VERIFY: Blocked agent moves to next chain. Blocking agent prioritizes the needed chain. SEVERITY: YELLOW
DETECT: Two agents are working on the same bug or making the same change.
MESSAGE to lower-priority agent:
STOP. [OTHER_AGENT] is already working on [TASK_DESCRIPTION].
This chain overlaps with their Chain [M]. Your work is redundant.
Revert your changes: git checkout -- [FILES]
Skip Chain [N] — mark it as "Duplicate of [OTHER_AGENT] Chain [M]" in your report.
Move to Chain [N+1].
VERIFY: One agent continues, the other moves on. No duplicate work. SEVERITY: YELLOW
DETECT: Agent terminal closed or session died. Branch has committed work.
MESSAGE (paste into new terminal in same worktree):
You are [CODENAME] agent RESUMING Sprint [XX].
Branch: sprint-[XX]/[AGENT_NAME]
Your previous session ended unexpectedly. Here's your state:
COMPLETED chains:
- Chain 1: COMPLETE — [SUMMARY]
- Chain 2: COMPLETE — [SUMMARY]
RESUME from: Chain [N] — [CHAIN_TITLE]
Last known state: [TRACING/FIXING/VERIFYING]
[DETAILS OF WHERE THEY LEFT OFF]
TERRITORY: [SAME AS ORIGINAL]
PROTOCOLS: [SAME AS ORIGINAL]
Continue from where you left off. Do not redo completed chains.
VERIFY: Agent picks up from the correct chain without redoing work. SEVERITY: YELLOW
DETECT: Agent terminal died AND had uncommitted changes that are now lost.
MESSAGE:
You are [CODENAME] agent RESTARTING Sprint [XX].
Branch: sprint-[XX]/[AGENT_NAME]
Your previous session ended and uncommitted work was lost.
COMPLETED chains (already committed):
- Chain 1: COMPLETE — [SUMMARY]
REDO Chain [N]: [CHAIN_TITLE]
Your previous attempt was [DESCRIPTION_OF_WHAT_WAS_LOST].
[INCLUDE THE ORIGINAL CHAIN DETAILS]
TERRITORY: [SAME AS ORIGINAL]
PROTOCOLS: [SAME AS ORIGINAL]
VERIFY: Agent redoes the lost work. Build + tests pass. SEVERITY: RED
DETECT: Agent is running but not responding to messages. Appears hung.
MESSAGE:
[First, kill the agent process and start fresh]
[Paste INT-16 or INT-17 recovery prompt depending on commit state]
Pre-restart steps:
- Check committed state:
git -C /path/to/worktree log --oneline -5 - Check uncommitted state:
git -C /path/to/worktree diff --stat - If uncommitted work exists, try to commit it:
git -C /path/to/worktree add . && git -C /path/to/worktree commit -m "WIP: auto-save before restart" - Kill and restart the agent tool
- Use INT-16 (if work was saved) or INT-17 (if work was lost)
VERIFY: New session starts cleanly. Agent resumes correct chain. SEVERITY: RED
DETECT: Agent's fix addresses part of the chain but misses edge cases or related code paths.
MESSAGE:
Your fix for Chain [N] is partially correct but incomplete.
What you fixed: [CORRECT_PART]
What you missed: [MISSING_PART]
Specifically:
- [MISSING_EDGE_CASE_1]
- [MISSING_CODE_PATH_2]
- [MISSING_VALIDATION_3]
Complete the fix before marking Chain [N] as done.
VERIFY: Agent extends the fix to cover missing cases. Full verification passes. SEVERITY: GREEN
DETECT: Agent's change fixes the original issue but introduces a new bug (tests fail, new error appears).
MESSAGE:
Your fix for Chain [N] introduced a regression.
Original issue: [FIXED ✓]
New issue: [DESCRIPTION OF NEW BUG]
Evidence: [TEST FAILURE / ERROR MESSAGE]
The problem is likely at [FILE:LINE]: [GUIDANCE]
Fix the regression without reverting the original fix.
Then re-verify both the original chain and the new issue.
VERIFY: Original fix still works AND regression is resolved. All tests pass. SEVERITY: YELLOW
DETECT: Agent added unnecessary abstraction, refactoring, or changes beyond the chain's scope.
MESSAGE:
Your fix for Chain [N] is correct but over-engineered.
The chain required: [MINIMAL_FIX_DESCRIPTION]
You also added: [UNNECESSARY_ADDITIONS]
Revert the unnecessary changes:
git checkout -- [FILES_WITH_UNNECESSARY_CHANGES]
Keep only the minimal fix. This sprint isn't about refactoring.
Document suggested improvements in your completion report under
"Suggested follow-up" for a future sprint.
VERIFY: Only the minimal fix remains. No unnecessary abstractions added. SEVERITY: GREEN
DETECT: Agent is implementing features or improvements not in their assigned chains.
MESSAGE:
STOP. You're adding [FEATURE/IMPROVEMENT] which is not in your chains.
Your assigned chains: [LIST]
What you're working on: [UNASSIGNED_WORK]
Revert: git checkout -- [FILES]
Return to your assigned chains. If you believe this improvement is
valuable, document it in your completion report under "Suggested
follow-up" — don't implement it now.
VERIFY: Agent reverts unassigned work and returns to their chains. SEVERITY: YELLOW
DETECT: Agent installs new packages/dependencies not justified by their chains.
MESSAGE:
HOLD. You added [DEPENDENCY] but your chains don't require it.
Revert: [COMMAND TO REMOVE DEPENDENCY]
(e.g., npm uninstall [package], go mod tidy, pip uninstall [package])
If you believe this dependency is necessary for your chain, explain
why in your completion report. Otherwise, use existing tools.
VERIFY: Dependency is removed. [dependency file] matches pre-sprint state for unrelated entries.
SEVERITY: YELLOW
DETECT: Agent edits shared config (package.json, go.mod, Makefile, .env.example) without justification.
MESSAGE:
HOLD. [CONFIG_FILE] is a shared configuration file.
Changes to shared configs must be justified and documented.
What did you change? Why was it necessary for your chain?
If justified: document the change and reasoning in your completion report.
If not justified: git checkout -- [CONFIG_FILE]
Note: LEAD has final authority on shared config changes during merge.
VERIFY: Change is either justified (documented) or reverted. SEVERITY: YELLOW
| ID | Category | Intervention | Severity |
|---|---|---|---|
| INT-01 | Territory | Wrong file modified | YELLOW |
| INT-02 | Territory | File created in wrong dir | YELLOW |
| INT-03 | Territory | Report doesn't match changes | GREEN |
| INT-04 | Chain | Chains out of order | YELLOW |
| INT-05 | Chain | No verification | YELLOW |
| INT-06 | Chain | Context-switching mid-chain | YELLOW |
| INT-07 | Stuck | Wrong root cause | YELLOW |
| INT-08 | Stuck | Needs information | GREEN |
| INT-09 | Stuck | Fundamentally blocked | YELLOW |
| INT-10 | P0 | Fix now (same agent) | RED |
| INT-11 | P0 | Route to different agent | RED |
| INT-12 | P0 | Stop sprint | RED |
| INT-13 | Conflict | Merge conflict | YELLOW |
| INT-14 | Conflict | Cross-agent dependency | YELLOW |
| INT-15 | Conflict | Duplicate work | YELLOW |
| INT-16 | Crash | Resume (committed work) | YELLOW |
| INT-17 | Crash | Restart (lost work) | RED |
| INT-18 | Crash | Force restart (hung) | RED |
| INT-19 | Quality | Incomplete fix | GREEN |
| INT-20 | Quality | New bug introduced | YELLOW |
| INT-21 | Quality | Over-engineered | GREEN |
| INT-22 | Scope | Unrequested features | YELLOW |
| INT-23 | Scope | Unnecessary dependencies | YELLOW |
| INT-24 | Scope | Shared config modified | YELLOW |
Related Documents:
- REACTIONS.md — Decision trees for runtime events
- STATUS-TRACKING.md — Sprint monitoring methodology
- OPERATORS-GUIDE.md — Full operator tutorial
- TEMPLATE-ACTIVATION.md — Activation prompt templates