Codons finished without actual last step & wrong path output

[nlavee]

🐛 Bug Report

Summary

1. Steps to Reproduce

1. Run demo-hank, pointing to a dir: npx hankweave ./hank.json <PATH_TO_DIR>
2. Let demo-hank runs through all codons, we focus on codon 3 build-world Assistant output.
3. Scroll down to output and I see the results output.

2. Expected Behavior

1. For codon 3 build-world, we see the following in Asisstant log:

[5:19:39 PM] Assistant
  │ Done! The world is built. Here's what I did:
  │
  │ ## Assembly Complete
  │
  │ 1. **Merged the data**: Combined `analysis-summary.json`, `world.json`, and the sentinel observations from `frognu-musings.md` into a single coherent `worldData.json` object.
  │
  │ 2. **Overlaid creative fields**:
  │    - `personality.description` → Frognu's characterization of your Early Bird nature
  │    - `personality.evidence` → Three specific observations about your download patterns
  │    - `busiestDay.funComment` → Frognu's poetic reflection on January 1, 1970
  │
  │ 3. **Enriched dialogue**: Added 1-2 bonus observations from the sentinel to each zone, capturing the deeper thoughts Frognu had while observing:
  │    - **Entrance**: Added context about downloading as "access points"
  │    - **Workspace**: Added the reflection about these shortcuts being maintained over years
  │    - **Secret**: Added the insight that the Google OAuth secret is the "actual work"
  │    - **Reflection**: Added the observation about intentional deletion and saying "no" to clutter
  │
  │ 4. **Merged with template**: Ran the merge script to combine `worldData.json` with the HTML template, injecting all data safely.
  │
  │ 5. **Verified output**: `output/downloads-wrapped.html` is 175 KB, fully populated, and ready to view.
  │
  │ The final HTML file should now be opening in your browser. Frognu's journey through your 25 carefully curated files is ready to tell your story.

[5:19:39 PM] Token Usage
  → Input: 18, Output: 5875
....

Expectation: HTML file should be opening in browser.

2. Run completes and we see infromation related to run:

┌─────────────────────────────────────────────────────────────────────────────────────┐
│                                   ✓ Run Complete                                    │
├─────────────────────────────────────────────────────────────────────────────────────┤
│ All 3 codons completed successfully in 4m 8s.                                       │
│                                                                                     │
│ Total cost:  $XXX│
│ Workspace:   /home/nlavee/.hankweave-executions/1771060481113-f9lh-1b394e/agentRoot │
│                                                                                     │
│ Look here for output files:                                                         │
│ /home/nlavee/.hankweave-executions/1771060481113-f9lh-1b394e/outputs                │
└─────────────────────────────────────────────────────────────────────────────────────┘

Expectation: /home/nlavee/.hankweave-executions/1771060481113-f9lh-1b394e/outputs exists

3. Actual Behavior

For 1, no HTML pops up in browse. Unsure whether this might be a function of running this on WSL, which might restrict web browser opening.

For 2, we actually don't have this path. Instead, we have the actual output downloads_wrapped.html in this path /home/nlavee/.hankweave-executions/1771060481113-f9lh-1b394e/**agentRoot**/output

6. Environment Details

• Operating System: Linux Ubuntu 24.04 (WSL)
• Project Name: Downloads Wrapped (from hank.json)
• Hankweave Version: 1.0.0 (from hank.json)
• Runtime: bunx / npx (as per README.md and prompts)
• Models Used: anthropic/claude-haiku (primary), anthropic/claude-haiku-4-5 (sentinel)
• Working Directory: /home/nlavee/demo-hank
• Execution Root (Example): ~/.hankweave-executions/1771060481113-f9lh-1b394e/

7. Logs & Stack Traces

model-validation.log seems to be just registering models, which we see:

[2026-02-14T09:15:45.115Z] [INFO] Self-test PASSED: All checks passed
[2026-02-14T09:15:45.115Z] [INFO]   - sdk_installed: ✓ Claude Agent SDK found (version installed)
[2026-02-14T09:15:45.115Z] [INFO]   - claude_cli_executable: ✓ Claude CLI detected at: /home/nlavee/.local/bin/claude
[2026-02-14T09:15:45.115Z] [INFO]   - authentication: ✓ Authentication configured via ANTHROPIC_API_KEY
[2026-02-14T09:15:45.115Z] [INFO] Self-test PASSED for Claude Haiku 4.5 (latest)    at /lib/utils.js:42:15
    at processTicksAndRejections (node:internal/process/task_queues:95:5)

Unsure whether there are other agent logs I can look at to figure out what's going on. How can we debug this?