Skip to content

Approvals answer in five vocabularies: one word set across the sheet, the card, /permissions and the transcript #1301

Open

Description

@santoshkumarradha

Split from #1295 (closed as superseded).

The drift

The sheet row "ask before running" cycles prompt / allow / deny (internal/config/settings.go:513, the row at :1943). The approval card answers 1 allow once · 2 always, this command · 3 deny · esc later (internal/session/answers.go:218-222). /permissions says asks every time / every call / refused. The transcript says allowed / denied (internal/tui3/consent.go:440-445). The task settle row says ask / auto. The rules syntax says bash:prompt, read:allow, deny rm -rf *. The phrase "allow or deny", which people carry away from the card, exists nowhere in the repo: the paraphrase is itself evidence the surfaces do not share words. The row at rest reads "ask before running · prompt", a tautology whose other states read as negation and double negation.

Twelve rows also show one label in the panel and another in the registry: "tool exceptions" vs "tool approvals", "per day" vs "daily budget", "ask before spending" vs "per plan", and nine more (audit-notes/settings-inventory.md). A person told one name searches for the other.

The law (approved in #1295)

One word set everywhere, stated as consequences: ask every time / run without asking / never run. No row label contains one of its own answers. The card, /permissions and the transcript adopt the sheet's words; the registry and the panel show one label per row.

Evidence: the mismatch table with every file:line is section 9 of audit-notes/permission-wiring.md; the walkthrough is section 2 of #1295.

—

Co-Authored-By: codeaf agentfield-bot@users.noreply.github.com


Drafted with CodeAF · reviewed and owned by the author

Activity

  1. added
    featureWork that adds a capability; developers break it into tasks
    area:chatThe v3 surface a person sits in front of (internal/tui3)
    sev:papercutA wording, a hint, a small wrongness that costs a moment
    help wantedBigger than a first issue but still decision free; the approach is settled, run with it
    on Sep 20, 2026
  2. 7vignesh commented on Oct 9, 2026

    @7vignesh
    Contributor

    @santoshkumarradha
    I'd like to take this one. I've read audit-notes/permission-wiring.md §9 and audit-notes/settings-inventory.md §(b), and confirmed on current dev that all four surfaces still carry their own vocabularies and the §(b) label mismatches are all live. Before I touch pervasive person-facing strings I want to lock the decisions the approved law leaves open, so I ship the words you actually approved rather than my guess at them.

    1. The exact word set. The law is "ask every time / run without asking / never run". Mapping the stored prompt / allow / deny:

    • prompt → ask every time
    • allow → run without asking
    • deny → never run

    Confirm these three strings verbatim? And the stored values stay prompt/allow/deny (persisted config and the rules syntax bash:prompt depend on them) — only the display changes. Correct?

    2. The sheet row label. "ask before running · prompt" is the tautology the law forbids ("no row label contains one of its own answers"). What should the row be called? Options: running tools, when the model wants to run a tool, or something you'd prefer.

    3. The transcript words. consent.go's decisionWord draws allowed / denied on a settled row. Should those become ran / did not run (the consequence, matching the set), or stay as past-tense allowed/denied? The law says the transcript adopts the sheet's words, but the sheet's words are forward-looking ("run without asking") and a transcript row is past-tense.

    4. Scope — is §(b) in this issue? The title names four surfaces (sheet, card, /permissions, transcript). The body also lists the twelve panel-vs-registry label mismatches from settings-inventory.md §(b) ("tool approvals" vs "tool exceptions", "daily budget" vs "per day", etc.). Are those twelve part of #1301, or do they belong to #1303 (the IA re-tab)? If they're in, do I default every row to the registry's label, or decide per-row (some panel labels read better, e.g. "compact at" over "context fill")?

    Once these are settled I'll do it in one lane: the sheet (config/settings.go), the card (session/answers.go), /permissions (tui3/permissions.go), the transcript (tui3/consent.go), and the manual page that quotes the approval words, with the manual gates and tests. One caveat up front: session/answers.go's package test suite does not build on my Windows box (internal/seniordev/engine/steploop is //go:build !windows), so I'll verify the card wording by reading plus the structural manual gate rather than a native unit run, and I'll say so on the PR.

  3. 7vignesh commented on Nov 14, 2026

    @7vignesh
    Contributor

    @santoshkumarradha @AbirAbbas re-checked the scope above against current dev: all four surfaces still carry their own vocabularies (sheet prompt/allow/deny at config/settings.go:506,2060, card allow once / always, this command / deny at session/answers.go:202-210, /permissions asks every time / every call / refused at tui3/permissions.go:75-77, transcript allowed/denied at tui3/consent.go:443), and the §(b) label mismatches are all live.

    Can I take this? I'm ready to open the lane on the four open questions above, defaulting to the verbatim law word set (prompt→ask every time, allow→run without asking, deny→never run) with stored values untouched, unless you'd rather settle the sheet row label and the transcript's past-tense wording first. Two additions to my file list since I posted: the sheet row's Hint string (config/settings.go:2061, it spells out the old words) and consentExpiredWord = "denied · no answer" (consent.go:197) both carry the vocabulary too. I'll hold for your call on the §(b) boundary vs #1303.

  4. AbirAbbas commented on Nov 16, 2026

    @AbirAbbas
    Collaborator

    @santoshkumarradha @AbirAbbas re-checked the scope above against current dev: all four surfaces still carry their own vocabularies (sheet prompt/allow/deny at config/settings.go:506,2060, card allow once / always, this command / deny at session/answers.go:202-210, /permissions asks every time / every call / refused at tui3/permissions.go:75-77, transcript allowed/denied at tui3/consent.go:443), and the §(b) label mismatches are all live.

    Can I take this? I'm ready to open the lane on the four open questions above, defaulting to the verbatim law word set (prompt→ask every time, allow→run without asking, deny→never run) with stored values untouched, unless you'd rather settle the sheet row label and the transcript's past-tense wording first. Two additions to my file list since I posted: the sheet row's Hint string (config/settings.go:2061, it spells out the old words) and consentExpiredWord = "denied · no answer" (consent.go:197) both carry the vocabulary too. I'll hold for your call on the §(b) boundary vs #1303.

    yup feel free to take it

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:chatThe v3 surface a person sits in front of (internal/tui3)featureWork that adds a capability; developers break it into taskshelp wantedBigger than a first issue but still decision free; the approach is settled, run with itsev:papercutA wording, a hint, a small wrongness that costs a moment

    Type

    No type

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions