← graph

tui-detail-view

TUI detail view: FLIP / eval scores not iteration-specific in cycles — stale scores show across iterations

Metadata

Statusdone
Assignedagent-953
Agent identityf51439356729d112a6c404803d88015d5b44832c6c584c62b96732b63c2b0c7e
Created2026-04-28T21:34:06.196716882+00:00
Started2026-04-28T21:51:41.196169093+00:00
Completed2026-04-28T22:14:09.740289868+00:00
Tagseval-scheduled
Tokens15103228 in / 25425 out
Eval score0.87
└ blocking impact0.88
└ completeness0.85
└ constraint fidelity0.85
└ coordination overhead0.88
└ correctness0.88
└ downstream usability0.85
└ efficiency0.78
└ intent fidelity0.78
└ style adherence0.92

Description

Description

When a task is in a cycle and iterates, the agency companion tasks (.flip-X, .evaluate-X, .assign-X) ALSO iterate alongside. But the TUI detail view shows FLIP/eval scores without labeling which iteration they came from — so iteration 2 of the user's task displays the iteration 1 FLIP score with no indication it's stale.

User quote: 'in the display in the tui detail view the flip is not iteration specific. but we are iterating the flip tasks too. all the tasks are iterating.'

Currently visible on this user's session via the create-agents-md ↔ verify-agents-md cycle (max-iterations=3): .flip-create-agents-md re-runs each iteration, but the detail panel just shows 'Score: 0.04 Source: flip' with one timestamp, no iteration label.

What to fix

  1. Each .flip-* / .evaluate-* run records or is tagged with the loop_iteration of the parent task at the time it ran.
  2. The TUI detail view's eval/flip section either:
    • Shows scores grouped by iteration (e.g. 'Iteration 1: flip 0.04 / eval 0.04 | Iteration 2: flip 0.65 / eval 0.74'), OR
    • Shows ONLY the current iteration's score and labels it as such ('Iteration 2: flip 0.65')
  3. The CLI wg show <task> output should match — same iteration labeling so users grepping logs aren't confused.

Likely files to touch

  • Wherever eval/flip records are stored on the task struct (probably src/graph.rs)
  • The TUI detail view renderer (likely src/tui/detail.rs or similar)
  • src/commands/show.rs (or wherever wg show formats evaluations)
  • The agency pipeline task — wherever .flip-* / .evaluate-* write their score back

Out of scope

  • Restyling the detail panel beyond what's needed to show iteration label
  • Changing how cycles iterate or how flip/eval are scheduled
  • Auto-archiving stale scores

Validation

  • Failing test first: a task with 2 completed iterations, each producing distinct flip scores, renders both in detail view (or current-only with explicit iteration label) — never just a single unlabeled score
  • Failing test: wg show <task> includes iteration label on each Evaluation entry
  • Manual smoke: on the existing create-agents-md cycle, after iteration 2 completes, detail view shows iteration 2's flip score (or both iterations clearly labeled)
  • No regression for non-cycle tasks (single iteration → no extra label noise, or label says 'Iteration 1' uniformly — pick one)
  • cargo build + cargo test pass

Depends on

Required by

Log