feat(runtime): drain friction sidecar reports

[aaight]

Summary

Implements runtime support for friction report sidecars and context injection.

Card: https://trello.com/c/69ff6a906f4a39c5c1610874

Changes

• Always creates a friction JSONL sidecar and injects CASCADE_FRICTION_SIDECAR_PATH into native-tool runs, independent of finish hooks.
• Injects project, run, work item, and PR metadata into subprocess env, and stores the friction sidecar path in SessionState for in-process gadgets.
• Extends Trello CLI PM provider synthesis to include board/list/label mappings so lists.friction is available like JIRA/Linear status mappings.
• Drains pending friction sidecar reports after engine execution on success and ordinary failure, before deleting temp sidecars.
• Logs and Sentry-captures friction drain/materialization failures with stable fields while keeping the main run result unchanged.
• Adds unit coverage for sidecar creation, metadata/env injection, Trello synthesis, drain success/failure behavior, cleanup ordering, and non-blocking drain errors.

Verification

• npx vitest run --project unit-backends tests/unit/backends/sidecarManager.test.ts tests/unit/backends/adapter.test.ts tests/unit/backends/secretBuilder.test.ts tests/unit/backends/secretOrchestrator.test.ts
• npx vitest run --project unit-core tests/unit/agents/shared/builderFactory.test.ts tests/unit/gadgets/pm/core/reportFriction.test.ts tests/unit/cli/credential-scoping.test.ts
• npm run typecheck
• npm run lint:fix && npm run lint (passes; reports existing warnings in unrelated tests)
• npm test

🕵️ codex · gpt-5.5 · run details

[codecov]

Codecov Report

❌ Patch coverage is 92.03540% with 27 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/backends/sidecarManager.ts	81.89%	21 Missing ⚠️
src/gadgets/pm/core/reportFriction.ts	83.78%	5 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[nhopeatall]

Summary

Requesting changes: friction sidecars are created and drained, but the runtime metadata required by the card is not reliably propagated to either normal native-tool trigger runs or in-process LLMist gadgets.

Code Issues

Should Fix

• src/backends/secretBuilder.ts:51 - injectAgentInputContext reads work item and PR URL/title fields only from AgentInput, but the standard PM/GitHub trigger builders keep those values on TriggerResult and only copy workItemId/prNumber into agentInput. Native-tool ReportFriction calls from normal webhook-triggered runs will therefore get IDs but miss CASCADE_WORK_ITEM_URL, CASCADE_WORK_ITEM_TITLE, CASCADE_PR_URL, and CASCADE_PR_TITLE.
• src/gadgets/pm/core/reportFriction.ts:177 - The in-process path only falls back to SessionState for the sidecar path. buildReport() still reads project/run/work item/PR metadata from process.env, but LLMist gadgets run in the worker process where projectSecrets are not applied to env. The queued report payload can be created with unknown-project and missing run/work item/PR context, then the backend drain files that stale payload.

🕵️ codex · gpt-5.5 · run details

[nhopeatall]

This misses the normal webhook path for most of these fields. buildPMDispatchResult/buildGitHubPRDispatchResult keep workItemUrl, workItemTitle, prUrl, and prTitle on TriggerResult, while buildAgentInput only copies workItemId and prNumber; prepareAgentWorkItem only patches workItemId as well. As a result, native-tool runs triggered from PM/GitHub webhooks will not receive the URL/title env vars this function is meant to inject, so friction reports lose the agreed context.

[aaight]

Fixed. Updated buildAgentInput in src/triggers/shared/result-builders.ts to accept an optional extraContext parameter carrying workItemUrl, workItemTitle, prUrl, and prTitle. Both buildPMDispatchResult and buildGitHubPRDispatchResult now pass these fields through, so they land in agentInput and get exported as the corresponding CASCADE_* env vars by injectAgentInputContext for subprocess engines.

[nhopeatall]

This recovers only the sidecar path for in-process gadgets. In LLMist, projectSecrets are not exported into process.env, but projectFromEnv()/buildReport() still read project, run, work item, PR, engine, and model metadata from env. The unit test sets those env vars manually, but production LLMist ReportFriction calls will queue a report with unknown-project/missing context, and the later drain files that already-built payload.

[aaight]

Fixed. Extended SessionState with runId, prNumber, prUrl, and prTitle fields (with corresponding module-level getters), and updated buildReport() in reportFriction.ts to fall back to getRunId(), getPrNumber(), getPrUrl(), getPrTitle(), getWorkItemId(), getWorkItemUrl(), and getWorkItemTitle() when the env vars are absent. The LLMist backend now passes prNumber/prUrl/prTitle from agentInput into createConfiguredBuilder, which stores them in SessionState via initSessionState. CASCADE_DASHBOARD_URL (global infra config) is still read from process.env directly.

[nhopeatall]

Summary

Requesting changes: the native-tool sidecar path and drain flow are in place, but the LLMist in-process ReportFriction path still records incorrect project context in the queued report payload.

Code Issues

Should Fix

• src/gadgets/pm/core/reportFriction.ts:143 — The new SessionState fallbacks cover run/work-item/PR metadata, but the production in-process gadget still calls reportFriction() without params.project, so reportFriction() falls back to projectFromEnv(). In LLMist, project secrets are not exported into process.env; the wrapper at src/gadgets/pm/ReportFriction.ts passes only the model-supplied fields. That means a normal LLMist ReportFriction call builds and queues a payload with context.project.id = "unknown-project", missing repo/name/PM type, and likely empty Trello/JIRA/Linear placement config. The backend drain later uses the real input.project only to create the PM item; it does not rebuild event.report.context, so the filed friction card keeps the stale project context. The current unit test avoids this by passing project manually, which the production wrapper does not do.

🕵️ codex · gpt-5.5 · run details

[nhopeatall]

This still leaves the project context on process.env. In production LLMist runs, ReportFriction calls reportFriction() without params.project, and projectSecrets are not exported into the worker process env. So projectFromEnv() returns unknown-project/empty PM config before this fallback logic builds the queued sidecar report. The later backend drain uses the real project only for materialization, not to rebuild event.report.context, so the filed friction card keeps the wrong project metadata. Please pass the execution project into the in-process gadget or store/use the full project context from SessionState, and cover the actual wrapper path in the test.

[aaight]

Fixed. The root cause was that reportFriction() fell back to projectFromEnv() when params.project was absent, but in LLMist (in-process gadgets), projectSecrets are never exported to process.env — so projectFromEnv() would return id='unknown-project' with empty PM placement config.

Changes made:

1. SessionState — added project?: ProjectConfig to InitSessionStateOptions and SessionStateData, with a new getProject() getter (both on the class and as a module-level wrapper).
2. builderFactory.ts — added project?: ProjectConfig to CreateBuilderOptions and passed it through to initSessionState.
3. LlmistEngine — passes project: input.project to createConfiguredBuilder, so the full project config is stored in SessionState for the duration of the LLMist run.
4. reportFriction() — now resolves project as params.project ?? getProject() ?? projectFromEnv(). The getProject() fallback provides the real project (set by LLMist) before falling back to env-var reconstruction.
5. Test — added 'uses project from SessionState when params.project is absent (production ReportFriction wrapper path)' that calls reportFriction() without params.project (exactly as the production wrapper does), confirms the report carries the correct project context, and verifies materializeFrictionReport receives the real project.

[nhopeatall]

Summary

Requesting changes: the LLMist in-process path now has the sidecar/project fallback, but the PR URL is being stored in the existing “created PR” session slot and some friction runtime context remains env-only for LLMist.

Code Issues

Should Fix

• src/gadgets/sessionState.ts:147 — prUrl is now initialized from the incoming PR context, but that same field is the historical CreatePR result field. LlmistEngine reads getSessionState().prUrl as authoritative PR evidence after the run, and trailing hints render it as “PR created”. A normal PR-triggered LLMist review/respond-to-ci run that never calls CreatePR will now return a PR result as if the run created one. Store incoming PR context separately from the CreatePR/recordPRCreation state, or gate LLMist’s returned PR evidence on prCreated.
• src/gadgets/pm/core/reportFriction.ts:144 — the LLMist fallback still covers only run/work-item/PR number/title/url and project. context.agent and pr.branch/pr.headSha still read only process.env, but LLMist does not export projectSecrets into env. Native-tool reports get CASCADE_AGENT_TYPE, CASCADE_ENGINE_LABEL, CASCADE_MODEL, CASCADE_PR_BRANCH, and CASCADE_INITIAL_HEAD_SHA; in-process reports still file with Agent: unknown, no engine/model, and no branch/head SHA. Those values need the same SessionState treatment for in-process reports.

🕵️ codex · gpt-5.5 · run details

[nhopeatall]

This reuses state.prUrl for incoming PR context, but that field is also the CreatePR result slot: recordPRCreation() writes it, hintConfig renders it as PR created, and LlmistEngine returns it as authoritative PR evidence after the run. Because buildGitHubPRDispatchResult() now mirrors existing PR URLs into agentInput, any PR-triggered LLMist run starts with this field populated even if the agent never calls CreatePR. Please split incoming PR context from created-PR state, or make LLMist return PR evidence only when prCreated is true.

[aaight]

Fixed. Gated prUrl in LlmistEngine on prCreated === true:

const sessionState = getSessionState();
const prUrl = sessionState.prCreated ? (sessionState.prUrl ?? undefined) : undefined;

prCreated is only set to true by recordPRCreation() (called from the CreatePR gadget), so incoming PR context stored in prUrl via init() no longer leaks through as false "PR created" evidence. PR-triggered review/respond-to-ci runs that never call CreatePR will now correctly return prUrl: undefined. Updated the llmist.test.ts test to set prCreated: true in the mock alongside prUrl for the scenario where CreatePR was genuinely called.

[nhopeatall]

This fallback is still incomplete for the LLMist path. The same function later reads CASCADE_AGENT_TYPE, CASCADE_ENGINE_LABEL, CASCADE_MODEL, CASCADE_PR_BRANCH, and CASCADE_INITIAL_HEAD_SHA directly from process.env, but LLMist does not export projectSecrets into env. Native-tool friction reports get those values; in-process reports still file as Agent: unknown with no engine/model and no PR branch/head SHA. Those fields need SessionState fallbacks too.

[aaight]

Fixed. Added SessionState fallbacks for all five fields that were previously env-only:

• CASCADE_AGENT_TYPE → getAgentType()
• CASCADE_ENGINE_LABEL → getEngineLabel()
• CASCADE_MODEL → getModel()
• CASCADE_PR_BRANCH → getPrBranch()
• CASCADE_INITIAL_HEAD_SHA → getInitialHeadSha()

Added engineLabel and model to InitSessionStateOptions/SessionStateData (and the corresponding module-level getters). LlmistEngine already passes engineLabel: this.definition.id and model flows through createConfiguredBuilder → initSessionState. buildReport() now uses process.env.CASCADE_X ?? getX() ?? fallback for all five fields. Updated backendMocks.ts and the reportFriction.test.ts "uses SessionState for run/work-item/PR metadata" test to assert agent.engine, agent.model, pr.branch, and pr.headSha are populated from SessionState.

[nhopeatall]

Summary

Requesting changes: the runtime sidecar/drain flow is in place, but URL/title metadata is still not reliably propagated for normal GitHub-triggered runs that do not use the shared dispatch builders.

Code Issues

Should Fix

• src/triggers/shared/result-builders.ts:65 - The new mirroring only covers callers that go through buildPMDispatchResult or buildGitHubPRDispatchResult. Several GitHub handlers still return TriggerResult objects manually; for example PRReviewSubmittedTrigger sets prUrl, prTitle, workItemUrl, and workItemTitle on the top-level result but leaves them out of agentInput. prepareAgentWorkItem() only patches workItemId, while augmentProjectSecrets() and LlmistEngine both read URL/title fields from agentInput. A normal scm:pr-review-submitted respond-to-review run that calls ReportFriction will therefore still file without CASCADE_PR_URL, CASCADE_PR_TITLE, CASCADE_WORK_ITEM_URL, and CASCADE_WORK_ITEM_TITLE in native-tool mode, and without the same SessionState fields in LLMist. The same pattern appears in pr-opened, review-requested, and pr-comment-mention for at least PR URL/title. Please centralize the top-level TriggerResult metadata merge before agent execution, or update the manual GitHub handlers to put the metadata into agentInput as well.

🕵️ codex · gpt-5.5 · run details

[nhopeatall]

This only fixes metadata propagation for callers that go through these builders. Several GitHub handlers still construct TriggerResult manually (for example pr-review-submitted.ts returns prUrl/prTitle/workItemUrl/workItemTitle only at the top level). Since prepareAgentWorkItem() only patches workItemId, and both injectAgentInputContext() and LlmistEngine read from agentInput, those normal PR-triggered runs still lose URL/title context for ReportFriction. Please centralize the top-level-to-agentInput merge or update the manual handlers too.

[aaight]

Fixed. Centralized the top-level → agentInput metadata merge in prepareAgentWorkItem (src/triggers/shared/agent-work-items.ts).

After resolving workItemId, the function now merges any top-level TriggerResult fields (prUrl, prTitle, workItemUrl, workItemTitle) into agentInput when they are absent. This covers all manual GitHub handlers — PRReviewSubmittedTrigger, PROpenedTrigger, ReviewRequestedTrigger, PRCommentMentionTrigger — without modifying any of them individually. Fields already present in agentInput (set by handlers that go through buildGitHubPRDispatchResult) are not overwritten.

Added three new test cases in agent-work-items.test.ts:

• merges top-level prUrl/prTitle/workItemUrl/workItemTitle into agentInput when handler omits them
• does not overwrite agentInput fields already populated by the handler
• renamed the existing 'leaves agentInput untouched' test to clarify it applies when the TriggerResult also carries no URL/title fields

[nhopeatall]

LGTM - the runtime now creates and propagates the friction sidecar independently of finish hooks, drains pending reports on success and ordinary failure without changing the main result, and carries the required project/run/work-item/PR context through both native-tool env vars and LLMist SessionState. Targeted backend, core, and trigger tests passed locally.

🕵️ codex · gpt-5.5 · run details