✨ Harness: descriptive plan ladder and push titles for viewers

[ibolton336]

Problem

The plan ladder the harness publishes to viewers (rendered in the console's session panel) has three rungs. Two of them were fixed text for every run: "Prepare workspace: clone, branch, grounding data" and "Agent works the stage task". Only the third named anything concrete (the push branch). The push tool calls viewers see as "Harness: git push (final)" did not say where they pushed. And the middle rung was one spinner for the entire agent turn, which can be minutes and hundreds of tool calls.

Change

Build the rungs from what the harness already has in hand, and keep the middle one moving.

First rung (commit 2): repository as host and path with credentials and .git dropped, the branch, and the number of analysis insights written. "no analysis insights" when the fetch returned none; nothing about insights when the run has no skills and never fetched them.

Prepare workspace: github.com/konveyor/coolstore on branch migration-1789401169, 49 analysis insights

Middle rung (commits 1 and 3): workflow stage position when there is one, the first non-empty line of the stage instructions (falling back to the agent prompt) stripped of Markdown markers and cut to 80 runes, the model, and the turn budget. Before the prompt is sent it reads up to 40 turns; once running, the ladder is re-emitted on every turn with progress:

Stage 2 of 3 — agent works the task: “Fix the findings from the assess stage.” (gemini-2.5-pro, turn 12 of 200)

Stage text is rendered with parameter values substituted, so the excerpt goes through the same redactor the closing-message path uses (#170). "Turn" is the measure the harness already reports as TurnsUsed and compares against the native limit in classifyOutcome: one per tool_call notification. The handoff prompt continues the count from where the primary stopped.

Push titles (commit 2): git push to branch <branch> (final) and (auto-commit watcher).

Plumbing (commit 3):

• SessionClient.SetTurnHandler fires from SendPrompt's loop each time the count advances; runStage re-emits the ladder from it.
• Re-emitting per turn would fill the tee's 32-frame replay ring and evict the push and outcome frames a late viewer needs, so the ring now keys plan frames: a new ladder replaces the previous one instead of stacking. Unkeyed frames accumulate as before.
• fetchAndWriteAnalysis returns the count it already logs. workflowStagePosition is split out of isIntermediateWorkflowStage so both read the stage metadata the same way; an index past the count now counts as invalid there too, which changes nothing for the intermediate check.

Verification

• go build, go vet, gofmt -l clean in harness/; full go test ./... in harness/ passes.
• New tests: the rung text across standalone/staged runs, prompt fallback, truncation, empty case, invalid stage metadata, secret masking, and progress wording with and without a budget; the prep rung's credential stripping, .git trimming, zero vs. unfetched insights and unparseable URL; the turn handler firing once per tool_call with the running count and not for other updates (demux-server test); and the replay ring keeping only the latest plan while unkeyed frames accumulate.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: b082ea2e-5d8a-4921-96d7-e671acd552fb

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[djzager]

Re-reviewed the updates. The redaction issue is fixed: full instructions are redacted before excerpting, and the regression test plus the original reproduction pass. No further actionable findings.