Replace low-value 'Resuming work' comment with conditional investigation findings gate and simplify skill rubrics

[sebastientaggart]

Problem to Fix

The /start skill posts a low-value "Resuming work" comment to GitHub tickets every time an agent resumes an existing issue. The comment typically restates the ticket title and provides no new information. Meanwhile, valuable insights discovered during the agent's investigation are lost when the conversation ends — they're never captured on the ticket itself.

Additionally, both /start and /submit-for-review contain detailed rubrics and enumerations for things that agents already handle well through natural language understanding, adding unnecessary complexity to the skill prompts without improving outcomes.

Why it Matters

For PMs and engineers watching a ticket, the "Resuming work" comment is noise that trains people to ignore ticket notifications. Conversely, genuinely useful investigation findings (root cause corrections, project-wide gotchas, related side-effects) disappear with the chat session — exactly the kind of durable knowledge that should live on the ticket.

The over-specified rubrics add cognitive load when maintaining the skills and consume agent context window for guidance the agents don't need.

General Approach

Replace the unconditional "Resuming work" comment in /start Case B with a conditional investigation findings gate: if the agent discovers anything novel and durable during investigation, present findings to the user with a simple post or skip choice before continuing to code. If nothing noteworthy was found, skip silently.

Simplify three over-specified rubrics across both skills by trusting agent judgment:

• Collapse the intent classification blocks (duplicated in Case A and Case B) from 5-line enumerations to a single principle
• Remove the complexity scale definitions (trivial/moderate/significant/extensive) — the field labels are self-explanatory
• Remove the resolution comment writing rules in /submit-for-review where the template already demonstrates the desired style

Complexity

Verification / QA effort: moderate
Changes are to skill prompt text only — no code logic, no config, no infrastructure — but the investigation findings gate introduces a new conditional flow in Case B that should be tested with a real resume.

Acceptance Criteria

• The "Resuming work" comment is no longer posted unconditionally when resuming a ticket
• When the agent's investigation surfaces novel findings, they are presented to the user with a post / skip gate before coding begins
• If investigation reveals nothing beyond what the ticket already states, no comment is posted and no gate is shown
• Intent classification blocks in /start are reduced to a single-line principle (applied in both Case A and Case B)
• The complexity scale definition is removed from /start
• The resolution comment writing rules are removed from /submit-for-review (except the #N form rule)

[sebastientaggart]

Agent Implementation Notes

Files to change

• skills/start.md — primary changes
• skills/submit-for-review.md — minor rubric removal

Changes in skills/start.md

1. Replace "Resuming work" comment (Case B Step 3, ~lines 369-373)

Remove the unconditional gh issue comment block. In its place, add a new conditional step between the current Step 2 (gate) and Step 3 (branch checkout). The new step:

• After the user says "go", check whether the investigation (from Step 1/2) surfaced anything not already stated or implied by the issue body — a root cause correction, a related side-effect, a project-wide gotcha.
• If yes: present numbered findings with the preamble "The investigation surfaced the following:" and offer post / skip.
  • post → write findings to <tmpdir>/investigation_comment.md via file-writing tool, then gh issue comment <number> --body-file <tmpdir>/investigation_comment.md using an ## Investigation Findings heading.
  • skip → proceed silently.
• If no qualifying findings: skip silently, no gate shown.

This requires creating the temp directory earlier in Case B (currently only Case A creates one in Step 3a). Add mkdir -p /tmp/CodeCannon && mktemp -d /tmp/CodeCannon/XXXXXX to Case B as well.

2. Simplify intent classification (Case A ~lines 161-166, Case B ~lines 325-329)

Replace both 5-line enumeration blocks with a single line:

Proceed only on unconditional approval. If the user's response includes conditions, questions, or adjustments, treat it as discussion — address their input and re-ask.

Keep the "Abandonment" line as a separate bullet since "stop means stop" is a hard behavioral rule, not judgment.

3. Remove complexity scale guidance (~lines 222-226)

Delete the 4-line rubric block. The <trivial | moderate | significant | extensive> enum in the field template is sufficient.

Changes in skills/submit-for-review.md

4. Remove resolution writing rules (~lines 313-317)

Delete the 4-bullet "Resolution writing rules" section. The template at lines 299-304 already specifies "written in plain language for a non-technical audience — no code, no file paths, no jargon." Move the #N form rule up into the template comment or the preceding paragraph since it's a mechanical requirement that can't be inferred.

Key decisions

• The investigation findings gate uses --body-file (same pattern as Case A's implementation notes) to avoid permission prompt issues.
• The litmus test for "qualifying findings" is deliberately left to agent judgment rather than enumerated — one sentence of guidance, not a rubric.
• Case B step numbering will shift: current Step 3 (branch checkout) becomes Step 4, current Step 4 (write code) becomes Step 5. The new findings gate becomes Step 3.

[sebastientaggart]

Resolution

Agents no longer post a generic "Resuming work" comment when resuming a ticket. Instead, if the investigation surfaces noteworthy findings, the agent presents them and asks whether to post or skip — if nothing new was found, no comment is posted at all. Several verbose rubrics across the start and submit-for-review skills were also simplified to trust agent judgment rather than spelling out what agents already understand.

See #149 for full technical details.