CtrlK
BlogDocsLog inGet started
Tessl Logo

benpiper-workspace/planning-execution-harness

Break down goals into multiple tasks and coordinate execution with gates and recovery. Based on Claw Code's agentic harness.

92

1.09x
Quality

90%

Does it follow best practices?

Impact

100%

1.09x

Average score across 3 eval scenarios

SecuritybySnyk

Passed

No known issues

Overview
Quality
Evals
Security
Files

EXAMPLES.md

Detailed Examples

Example 1: Resume Optimization

Goal: "Optimize my resume for 3 software engineer job postings"

Plan:

Task 1: Analyze the 3 job postings and extract top 10 skills (parallel)
Task 2: Review current resume and identify gaps (depends on Task 1)
Task 3: Rewrite experience section to emphasize required skills (depends on Task 2)
Task 4: Add skills section with keywords (depends on Task 1)
Task 5: Proofread and format (depends on Task 3, Task 4)

User Feedback: "Approved, but skip the skills section — I'll add that manually"

Updated Plan:

Task 1: Analyze job postings and extract top 10 skills
Task 2: Review current resume and identify gaps (depends on Task 1)
Task 3: Rewrite experience section to emphasize required skills (depends on Task 2)
Task 4: Proofread and format (depends on Task 3)

Execution:

[Task 1/4] ✓ Analyzed 3 job postings
  Top skills: Python (all 3), AWS (2/3), System Design (2/3), Leadership (all 3), React (2/3)

[Task 2/4] ✓ Reviewed resume gaps
  Have: Python, Leadership
  Missing: AWS, System Design deep dive, React specifics

[Task 3/4] ✓ Rewrote experience section
  Added AWS/EC2 project details
  Expanded system design examples
  Mentioned React contributions

[Task 4/4] ✓ Proofread and formatted
  Fixed typo in "architected"
  Aligned bullet points

Outcome: ✓ Resume optimized. Ready to submit.


Example 2: Debug with Recovery

Goal: "Fix why my API keeps timing out on large requests"

Plan:

Task 1: Check request size limits
Task 2: Check database query performance
Task 3: Check network timeout settings
Task 4: Monitor actual timeout location if still failing

User: "Approved"

Execution:

[Task 1/4] ✓ Checked request size limits
  Limit: 100MB
  Actual large request: 50MB
  Not the issue.

[Task 2/4] ✓ Checked database query
  Found: N+1 query bug
  Query takes 45 seconds on large dataset
  Recovery needed.

  Recovery attempt: Add query optimization with joins
  [Task 2/4 RETRY] ✓ Query now takes 2 seconds

[Task 3/4] ✓ Checked timeout settings
  Timeout: 30 seconds
  With optimized query: well within limit

[Task 4/4] — Skipped (no longer needed, issue was Query optimization)

Outcome: ✓ Found N+1 query bug. Optimized. API now handles large requests in <5 seconds.


Example 3: Multi-Step Deployment

Goal: "Deploy new authentication system to production"

Plan:

Task 1: Run full test suite (no side effects)
Task 2: Review deployment plan with team (approval gate)
Task 3: Deploy to staging (test before prod)
Task 4: Run smoke tests on staging
Task 5: Deploy to production (irreversible)
Task 6: Monitor production for 1 hour

User Review (Task 2 gate):

  • "Approved, but add a rollback plan"

Updated Plan:

Task 1: Run full test suite
Task 2: Create rollback procedure and document it
Task 3: Review deployment with team (approval gate)
Task 4: Deploy to staging
Task 5: Run smoke tests on staging
Task 6: Deploy to production
Task 7: Monitor for 1 hour and rollback if needed

Execution:

[Task 1/7] ✓ Full test suite passing (1200 tests)

[Task 2/7] ✓ Created rollback procedure
  Documented: database migration rollback, service restart order, config revert

[Task 3/7] GATE: Team review requested
  Team approves deployment plan at 2:00 PM

[Task 4/7] ✓ Deployed to staging

[Task 5/7] ✓ Smoke tests passed
  Login works, token validation works, SSO works

[Task 6/7] ✓ Deployed to production

[Task 7/7] ✗ Production monitoring: 5% login failure rate detected at 2:15 PM
  
  Recovery: ROLLBACK (prepared in Task 2)
  [Task 7/7 RECOVERY] ✓ Rolled back to previous version
  Login failure rate drops to 0.1% (baseline)
  
  Analysis: New SSO integration failing for 5% of users
  Escalated to auth team for investigation

Outcome: ⚠ Partial success. Main system working. SSO integration needs fix. Safely rolled back.


Example 4: Writing Code with Refinement

Goal: "Write a Python function that detects typos in resume text"

Plan:

Task 1: Define typo types we want to catch (spell check, double spaces, punctuation)
Task 2: Write function skeleton with docstring
Task 3: Implement spell check using dictionary
Task 4: Implement double-space detection
Task 5: Implement punctuation checks
Task 6: Write unit tests
Task 7: Test against sample resumes

User: "This looks good, but add support for common business jargon so we don't flag 'synergize' as a typo"

Updated Task 1:

Task 1: Define typo types AND approved jargon list

Execution:

[Task 1/7] ✓ Defined typo types and jargon whitelist
  Typos: misspellings, double spaces, incorrect punctuation
  Jargon: synergize, leverage, ideate, pivot, etc.

[Task 2/7] ✓ Function skeleton written

[Task 3/7] ✓ Spell check implemented

[Task 4/7] ✓ Double-space detection implemented

[Task 5/7] ✓ Punctuation checks implemented

[Task 6/7] ✓ Unit tests written (15 tests)

[Task 7/7] ✓ Tested against 5 sample resumes
  False positive rate: 2% (acceptable)
  Caught real typos: 92% detection rate

Outcome: ✓ Function complete and tested.


Edge Cases

Edge Case 1: Plan Rejected Entirely

User Plan Review: "No, this approach is wrong. Try a completely different strategy."

Response: Return to planning stage. Generate new plan. Repeat until approved.

Edge Case 2: Task Dependencies Change Mid-Execution

During Execution: Task 3 fails, but it wasn't needed for Task 4 anyway.

Recovery: Ask user: "Skip Task 3 and continue with Task 4?"

Wait for approval before changing the plan mid-execution.

Edge Case 3: Transient Error Retry Loop

Situation: Network timeout, retry, timeout, retry, timeout...

Recovery Policy:

  • Retry transient errors max 2 times
  • Then escalate to user: "Network unreliable. Abort, or try again manually?"

Don't retry forever.

Edge Case 4: User Changes Requirements Mid-Execution

During Execution: "Actually, skip Task 4 and jump to Task 5"

Response:

  1. Acknowledge the change
  2. Ask: "Confirm: skip Task 4, proceed to Task 5?"
  3. Update the log to show the change was user-requested
  4. Continue execution

Key Takeaways

  • Decomposition is critical — Bad plans lead to bad execution
  • Gates prevent disasters — Always show the plan before acting
  • Recovery is strategic — Each failure type needs its own response
  • Logging is observability — You can only fix what you can see
  • Flexibility matters — Plans change. Be ready to adapt.

EXAMPLES.md

IMPLEMENTATION.md

PROMPT.md

PUBLISHING.md

QUICKSTART.md

README.md

REFERENCES.md

SKILL.md

SPEC.md

tile.json