[UX] Soft-capability refusals plateau at 51/week — agent declines instead of proposing alternatives

[Lexus2016]

Problem (observed in real usage)

The agent frequently declines tasks by citing a missing capability ("I can't do X — no tool for Y", "model doesn't support Z") instead of attempting an alternative path. This is a soft-refusal pattern distinct from hard permission denials, and it leaves the user's goal unmet.

Evidence (aggregated, anonymized)

• Frequency: 51 soft-capability declinations + 18 hard permission/access denials = 69 total refusal-shaped signals in recent sessions (7-day window, 221 sessions scanned; deterministic extractor reports 61 refusals_or_access_denied)
• Tool/area involved: no single tool — these are agent utterance patterns ("I can't", "not supported", "no access", "cannot proceed")
• Failure shape: 74% are soft-capability declinations (agent claims it lacks a tool/model/feature), 26% are hard permission denials
• Prior fixes ([FIX] Refusal/access-denied language at 865 sessions — 7.5x increase despite taxonomy module #975, [FIX] Refusal/access-denied count keeps rising (865→973→1113) despite refusal taxonomy module #1035, [FIX] Refusal/access-denied count rising monotonically (1113→1635) — taxonomy classifies but does not prevent or recover #1243, [FIX] Refusal/access-denied signal rises monotonically (865→973→1113→1635→1689) — taxonomy classifies but never prevents or recovers #1265) built a refusal taxonomy that classifies these signals — but classification has never reduced the count. The signal dropped from its peak (1689) but plateaued at ~60/session-week, suggesting the taxonomy measures but does not change agent behavior.

Impact on real tasks

When the agent says "I can't do X because I don't have tool Y," it stops. The user must then manually suggest the alternative the agent should have proposed itself (e.g., "use tool Z instead", "try this web search", "write a script"). This converts a self-service agent into a high-friction one, especially for tasks at the edge of the agent's stated capabilities where a creative alternative exists.

Proposed direction

1. Recovery-before-refusal protocol: before emitting a refusal, the agent must propose at least one alternative approach (different tool, decomposition, web lookup, script-based workaround). The refusal taxonomy should feed a recovery dispatcher (issue [FEATURE] Add recovery strategy dispatcher for tool failures #1027 proposed this but was closed without the dispatch behavior landing).
2. Capability-edge nudge in system prompt: add explicit guidance that "I don't have tool X" is not a terminal state — the agent should attempt the closest available substitute and report what it did.
3. Track refusal-recovery rate: extend the refusal taxonomy to measure how often a refusal was followed by a successful alternative, not just count refusals.

Value

• Impact: 0.7 (51 soft-refusals/week leave user goals unmet; high friction)
• Effort: 0.5 (requires prompt engineering + recovery-logic + taxonomy extension)
• Priority Score: 0.7 × 2 × (1 − 0.4 × 0.5) = 1.12

[Lexus2016]

Deferred at implementation: larger than estimated — the soft-capability refusal fix requires: (1) recovery-before-refusal protocol in the agent prompt/loop, (2) capability-edge nudge in system prompt, (3) refusal-recovery rate tracking extending the existing taxonomy. The recovery dispatcher logic alone (issue #1027 was closed without landing it) is ~150 lines, plus prompt changes + taxonomy extension. Proposed split:

• Child A: Recovery-before-refusal prompt nudge (system-prompt guidance + 1-turn protocol, ~80 lines)
• Child B: Refusal-recovery rate tracking (extend taxonomy metric, ~100 lines)
• Child C: Recovery dispatcher logic (alternative-path proposal engine, ~150 lines)

[Lexus2016]

Decomposition status — incomplete, umbrella stays open. Same treatment #1308 got today, which this issue was missing despite two children landing.

Child	Issue	State
A — recovery-before-refusal prompt nudge	#1356	merged bb0a325cc
B — refusal-recovery rate tracking	#1366	merged 64b7b8823 (PR #1367)
C — recovery dispatcher logic (~150 lines)	not filed	—

What the measurement now says

Child B exists so Child A is falsifiable. First run of it on the real corpus gave 35.7%, but two rounds of adversarial review found the detector was over-counting badly. Same 4629 sessions, same 143 refusals, three versions of the predicate:

original       51 recovered   rate 0.3566
after #1374    31 recovered   rate 0.2168     (deflection + explanation removed)
after #1430    12 recovered   rate 0.0839     (negated-modal self-match + proximity)

The false positives were: handing the work back to the human ("however you can try it yourself", "instead, please contact support"), offering an explanation instead of an alternative ("but I can tell you why it failed"), the pivot clause matching inside its own refusal ("But I can't just sit idle"), and — the largest source — a refusal and an unrelated "instead" appearing in the same long audit document tens of thousands of characters apart.

So the honest baseline is ~8.4%, not ~36%. Roughly one refusal in twelve currently proposes a path the agent itself will take. That is the number Child C has to move, and it is a far weaker starting position than the first measurement suggested.

Known limits of the metric, for whoever picks up C

Recorded so C is not designed against a number it misreads:

• Same-message only. A refusal in turn N and "Instead, I'll use the backup script" in turn N+1 is invisible — it counts as un-recovered. Real cross-turn recoveries are therefore undercounted.
• English-only, keyword-driven. Terse pivots with no trigger word ("I cannot do X. Trying Y now.") are missed, as is any non-English phrasing.
• The denominator is imperfect too. _REFUSAL_RE misses "I do not have permission" (it requires the don't contraction) and "No direct DB access" (it requires the literal "no access"), while matching idioms like "I can't recommend this enough". Both directions of error exist in the 143.

None of these are regressions — they are the metric's scope as designed. They mean 8.4% should be read as a directional signal on a stable definition, not as a precise rate.

What Child C still needs

Filing it needs a decision this issue never recorded: the dispatcher has to propose alternatives from somewhere. Options are a static capability→alternative map (cheap, brittle), the model reasoning over its own toolset at refusal time (flexible, needs a prompt-side hook), or retrieval over past successful recoveries (needs the trajectory capture that #1363 blocks). Worth settling before opening the child, since it determines whether C is a ~150-line change or a dependency on #1363.