Repository navigation
New-harness detection and setup nudges (surfaces 1–2) - #626
Merged
realtonyyoung merged 12 commits intoAug 20, 2026
Merged
Conversation
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Detects, after initial setup, that a supported coding harness is installed but kcap is not wired into it, and prompts to set it up. Implements the spec's PR 1 (independently shippable client surfaces); surfaces 3–4 (daemon/ingest inventory → server notification, and the server-side machine-went-quiet backstop) follow in PRs 2–3. Core substrate: - HarnessCatalog: single Core table binding vendor id + label + install flag + detection selector; the App's AgentVendors re-derives from it, and a conformance test pins it against VendorSelection.KnownVendorFlags. - HarnessIntegrationProbe.IsWired: "is kcap wired into vendor X?", the single source of truth now shared by the kcap status Hooks line (Claude's enabled-plugin and Codex's hooks-reference checks moved to Core; status delegates to them, behavior-preserving). - HarnessOfferLedger/Store: per-machine harness-offers-v1.json (atomic write, corrupt→empty) plus the shared 6h evaluation throttle stamp. - HarnessNudge.Nudgeable: pure predicate (detected ∧ !wired ∧ !declined ∧ past the 7-day re-offer floor). Surfaces: - 1: HarnessNudgeEmitter fragment wired into all 9 hook commands' SessionStart additionalContext, next to the work-items nudge. - 2: HarnessSetupNotice — exit-time interactive stderr notice (TTY-gated, human-facing commands only), mirroring the update-available notice and sharing surface 1's throttle. - kcap status: passive "installed but kcap not configured" line (ledger- independent — always tells the truth). - kcap harness list|dismiss|reset: inspect/silence/re-enable the nudges; all bypass the throttle. - kcap setup stamps offered harnesses so it never re-nudges a vendor just seen at setup; never writes/overwrites a dismissal. - Profile.DisableHarnessNudge (config key disable_harness_nudge) opts out surfaces 1–2 entirely. Tests: catalog conformance, predicate per-vendor, ledger/store (incl. dismissal-preservation), IsWired, emitter (fragment/notice/throttle/fold/ opt-out/exception→null). README + help + config help updated. Closes #625 AI-2118 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
PR Summary by QodoDetect unwired coding harnesses and nudge users to configure them
AI Description
Diagram
High-Level Assessment
Files changed (32)
|
Code Review by Qodo
1.
|
![qodo-code-review[bot]](https://avatars.githubusercontent.com/in/484649?s=60&v=4){kind=small}
…ness, catalog delegate - Offer ledger updates now serialize across processes via ConfigFileLock and write through a per-process-unique temp file, so a concurrent hook/setup/ command can no longer overwrite another's change (notably lose a dismissal). (qodo #4) - Load normalizes a null `vendors` member to an empty dictionary, honoring the corrupt-to-empty contract so management commands can't null-deref. (qodo #5) - Claude/Codex wired-checks moved into their own installers (ClaudePluginInstaller.IsPluginEnabled / CodexHooksInstaller.ReferencesKcapHook) and read via a new Core SharedFileText (FileShare.ReadWrite), so probing never blocks an agent writing its own settings.json/hooks.json on Windows. (qodo #2) - Wired-check folded into a per-entry delegate on HarnessCatalog (the single Core registration site); HarnessIntegrationProbe is now thin generic dispatch, no per-vendor switch in shared code. (qodo #1) - `harness` added to Program.cs offline-command allowlist, so `kcap harness list|dismiss|reset` works with no server configured. (qodo #7) - Trimmed predicate comments that restated the code. (qodo #3) - Added null-vendors ledger test. (qodo #6 — "setup stamping fails compilation" — is a false positive: `detected` is CodingAgentsStep.DetectedAgents whose members are bool, and CI's build passed.) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…osed throttle, honest dismiss, snapshot-consistent wiring - Offer-ledger Update takes a lock timeout; the hook/best-effort StampOffered path uses a short 1s wait so a contended ledger can't stall a SessionStart hook past a host's ~5s budget (management commands keep the 5s wait). (Codex P1 #1) - SharedFileText opens with FileShare.Delete too, so a concurrent read can't make the ledger's atomic File.Move-overwrite fail with a sharing violation on Windows (which would silently drop a dismissal). (Codex P1 #2) - TryClaimCheck now fails CLOSED: if the throttle stamp can't be written the same failure blocks the ledger stamp, so returning true would nudge on every hook — suppress instead. (Codex P1 #3) - Update returns whether the change persisted; `kcap harness dismiss|reset` surface a false as exit 1 instead of falsely reporting success (a lost dismissal would otherwise silently revive the nudge). (Codex P1 #4) - Kiro/Pi/OpenCode wired-probes now thread the injected KiroHome/PiAgentDir/ OpenCodeConfigDir overrides (new optional params on the *Paths helpers) so detection and wiring consume the same AgentDetectionInputs snapshot, per the Core API contract. (Codex P2 #5) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ck, pure snapshot wiring - On lock-acquire failure Update no longer does a lockless read-modify-write (which could recreate the lost-dismissal race); it returns false without mutating. Hook callers ignore it (best-effort); dismiss/reset already report false as exit 1. (Codex R2 P1 #1) - Hook-path stamping uses a zero-wait (non-blocking) lock try, so a SessionStart hook spends none of its exit-budget safety reserve waiting on the mutex; on contention it simply skips stamping. (Codex R2 P1 #2) - Kiro/Pi/OpenCode wired-probes resolve paths via new pure helpers (KcapAgentJsonPure / KcapExtensionPure / KcapPluginPure) built from the existing *Pure roots, so a null injected override means "unset → home default" and never re-reads ambient env — honoring AgentDetectionInputs' null-as-unset contract and keeping detection and wiring on the same snapshot. Reverted the round-1 non-pure override params (superseded). Added a Kiro override-threading test. (Codex R2 P2 #3) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- StampOffered takes a lockTimeout; the hook path passes TimeSpan.Zero while setup omits it and gets the normal serialized 5s wait, so a momentary concurrent writer no longer makes setup silently skip stamping the offered vendors (which would drop the 7-day floor). (Codex R3 P2 #1) - Copilot/Gemini/Antigravity wired-probes now use pure path helpers (KcapHooksJsonPure / SettingsJsonPure / GlobalHooksJsonPure) built from the existing *Pure roots, so every override the snapshot carries is resolved without an ambient env re-read. Claude/Codex necessarily read CLAUDE_CONFIG_DIR/CODEX_HOME from ambient because those roots aren't part of AgentDetectionInputs (detection is PATH-only for them) — documented as the sole, production-coincident exception on HarnessCatalog. (Codex R3 P2 #2) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
GlobalHooksJsonPure built from RootPure (<gemini>/antigravity, the data root), but production GlobalHooksJson resolves through GuiConfigRoot (<gemini>/config), so a normally-wired Antigravity install was probed at the wrong path and reported unwired — spurious nudges while `kcap status` said configured. Add GuiConfigRootPure (<gemini>/config) and build the pure hooks path from it, mirroring production. Add HarnessWiredPathParityTests pinning every pure wiring path to its production layout so a wrong root/segment fails a test. (Codex R4 P1) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…rness-detection-c9f55a # Conflicts: # src/Capacitor.Cli.Core/Harness/Kiro/KiroPaths.cs
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Detects, after initial
kcap setup, that a supported coding harness (Claude, Codex, Cursor, Copilot, Gemini, Kiro, Pi, OpenCode, Antigravity) is installed but kcap is not wired into it, and prompts the user to set it up — so a harness installed later (or switched to) starts being recorded without a manual re-run of setup.This is PR 1 of the spec's 3-PR slicing (the independently-shippable client surfaces). Surface 3 (daemon/hook-ingest harness inventory → Capacitor UI notification) and surface 4 (server-side machine-went-quiet backstop) follow in PR 2 (kcap-cli client fragment) and PR 3 (kcap-server). Design spec rides this PR:
docs/superpowers/specs/2026-08-19-new-harness-detection-nudge-design.md(reviewed clean by Copilot + Kiro spec-review flows).What's here
Core substrate
HarnessCatalog— the single Core table binding vendor id + label + install flag + detection selector. The desktop app'sAgentVendorsnow re-derives from it, and a conformance test pins it againstVendorSelection.KnownVendorFlagsso a tenth harness fails a test rather than silently missing every surface.HarnessIntegrationProbe.IsWired(vendorId, AgentDetectionInputs)— "is kcap wired into vendor X?", callable from CLI and (later) daemon. To avoidkcap statusever contradicting the nudge, Claude's enabled-plugin check and Codex's hooks-reference check moved to Core as the shared source of truth;StatusCommanddelegates to them (behavior-preserving).HarnessOfferLedger/HarnessOfferStore— per-machine~/.config/kcap/harness-offers-v1.json(atomic write, corrupt→empty) plus the shared 6-hour evaluation throttle stamp.HarnessNudge.Nudgeable— pure predicate: detected ∧ not wired ∧ not declined ∧ past the 7-day re-offer floor.Surfaces
HarnessNudgeEmitterfragment wired into all 9 hook commands next to the work-items nudge; tells the agent to offerkcap plugin install --<vendor>.HarnessSetupNotice, an exit-time one-liner (TTY-gated, human-facing commands only), mirroring the "update available" notice and sharing surface 1's throttle.kcap statusgains a passive "‹Vendor› installed but kcap not configured — run…" line (ledger-independent: status always tells the truth, even for dismissed vendors).kcap harness list | dismiss | resetto inspect / silence / re-enable the nudges (all bypass the throttle).kcap setupstamps the harnesses it offered, so the nudge never re-offers a vendor just seen at setup; it never writes or overwrites a dismissal.Profile.DisableHarnessNudge(kcap config set disable_harness_nudge true) turns off surfaces 1–2 entirely.Testing
IsWired; emitter (fragment / notice / throttle / multi-vendor fold / opt-out / exception→null).Harness*, CLIHarness.*namespace,SessionStartMemory, App Agents step). TheCommands.HarnessCursor tests that read the real~/.cursorare the pre-existing local env-coupling (they fail identically onmainon a machine with kcap set up for Cursor); under an isolated HOME they pass 42/42, matching CI.README,
help-usage.txt, andconfig sethelp updated in this PR.Closes #625
AI-2118
🤖 Generated with Claude Code