feat: capture agent-stated intent via MCP tool into session.jsonl #27

Activity

1. codeprakhar25

   added

   enhancementNew feature or requestNew feature or request

   on May 21, 2026
2. MHASK commented on Jul 17, 2026

   

   MHASK

   on Jul 17, 2026

   More actions

   This is a strong direction. I would treat agent-stated intent as useful review evidence, but not as the only source of truth. The agent has context, but it is also the actor being audited.

   A few schema/design suggestions that may make this more robust:

   • Store intent.source, e.g. agent_stated, human_declared, commit_message, issue_title, fallback_classifier. That lets reports distinguish intent captured from the agent from intent declared by a human.
   • Store intent.scope or intent_target, e.g. files, symbols, issue number, task id, or PR id the intent claims to address. This helps detect drift later.
   • Keep both agent_stated_intent and observed_change_summary. If they disagree, surface that mismatch instead of overwriting one with the other.
   • Add risk_categories[] separately from intent_type, e.g. auth, payments, ci, dependencies, infra, secrets, migrations. A change can be a refactor and still touch auth.
   • Consider a confidence or evidence field for derived values. Human-declared or issue-linked intent is different from a classifier guess.
   • Record the commit/session boundary carefully for multi-step sessions. A session-level intent can become misleading if one long agent session produces several commits with different purposes.

   For PR comments, the most useful output may be:

   Claimed intent: <agent/human stated intent>
   Observed change: <diff-derived summary>
   Risk categories: <auth, deps, ci, ...>
   Potential drift: <none / touched billing while intent mentions login>
   

   That makes intent capture actionable for maintainers rather than just another metadata field.

   ❤️React with ❤️1Reacted by Prakhar Khatri
3. codeprakhar25 commented on Jul 18, 2026

   

   codeprakhar25

   on Jul 18, 2026

   OwnerAuthor

   More actions

   Thanks, this is very useful!

   So the current status is: set_intent shipped in v0.1.29. intent + intent_type (bugfix/feature/refactor/test/docs/security/performance/config/dependency) flow through session.jsonl → prepare-ledger → finalize-ledger → traces → reports today.

   Taking two of your suggestions as follow-up work, since they're cheap and add real signal without another schema layer:

   • intent.source (agent_stated | fallback) — right now intent can come from a set_intent event or from pending-context payload with no marker distinguishing them. Small addition, closes that gap.
   • risk_categories[] — derivable deterministically from touched paths (auth/, payments/, .github/workflows/, Cargo.toml/package.json, migrations/) with no LLM call, consistent with the zero-cost design goal already in this issue.
   • Issue/ task ID can be a great addition here intent.scope/intent_target

   Already implemented/thoughts:

   • separate confidence/evidence field — redundant with source; the source tag already implies confidence tier.
   • observed_change_summary + drift detection against claimed intent — agree this is the highest-value idea here, but it's a real feature (diff summarizer + comparison), not a field. Worth its own issue rather than folding into this one.

   Session/commit boundary: already handled — finalize-ledger.py strips consumed type: intent events from session.jsonl right after the commit they belong to, so long sessions with multiple commits don't bleed intent across boundaries.

   The main challenge with more increasing this is how scalably and eff. we can store this in large codebase!
4. MHASK commented on Jul 18, 2026

   

   MHASK

   on Jul 18, 2026

   More actions

   That breakdown makes sense, especially keeping source as the confidence tier instead of adding another field.

   For large repos, I would probably avoid storing symbol/file-level derived data directly on every session event. A lighter shape could be:

   • keep the raw intent event small: intent, intent_type, source, scope/task ref
   • compute risk_categories from the touched paths during ledger/finalize
   • store a compact per-commit summary keyed by commit/session id
   • keep any expensive observed-change summary as a generated artifact, not mandatory ledger state

   So the durable ledger stays cheap, and richer drift detection can be opt-in or cached by commit hash. That also leaves room for large monorepos where the same touched-path classification may be enough for review routing without persisting a full semantic summary every time.

   👍React with 👍1Reacted by Prakhar Khatri
5. codeprakhar25 commented on Jul 28, 2026

   

   codeprakhar25

   on Jul 28, 2026

   OwnerAuthor

   More actions

   Got late, missed this!

   But yeah, this should be the path, opt-in for large repos, especially monorepos. Will double check for the scope/task-ref definition. Also the definition will be evolved as it is developed. Would you be open to contribute to the repo with this one?