Repository navigation
Conversation
`--policy`, `--min-confidence`, and `--input` for the policy commands, and one consumer added to the existing `--max-steps` declaration rather than a second flag under the same name. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The neutral half of the policy feature: the candidate projection a policy head chooses between, a content digest that detects an action which changed nothing, and the caller-supplied text map a field is filled from. The digest excludes refs on purpose. Refs are reissued per snapshot generation, so including them would make every screen look changed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
One provider behind a named abstraction, over global fetch with no new dependency. The credential comes from TYPESAFE_API_KEY in the environment only, so it never reaches argv, a recorded script, or an MCP tool schema. Every field of a response is validated against the candidate set rather than trusted, and an unreachable head fails with the status as a typed reason plus the documented fallback: the agent's own snapshot-and-choose loop. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Snapshot, decide, act, verify, repeat. Three device operations behind a narrow port so the loop is testable without a device. What the loop had to learn from a real iOS sign-in flow: - A settled observation only proves the local UI went quiet, so a transition waiting on the network still reads as the old screen. One grace re-read separates a slow transition from an action that did nothing. - A masked field reformats what it receives, so the runner reports TEXT_ENTRY_MISMATCH on text that arrived. The screen decides, not the write. - A digit-box code field never reflects its value at all. A confirmed value is sufficient evidence of a write but not necessary; a screen that moved on is the other half, and the on-screen keypad is the fallback when neither holds. - A press on a text field only focuses it, so the supplied entry rather than the field's emptiness decides whether to write. - A head asked whether progress is blocked says yes on any sign-in screen. A concrete target at 0.9 confidence outranks that flag, or the loop stalls on step one of every sign-in. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`suggest <goal>` prints one typed decision without acting. `act <goal>` runs the loop. Both are local-cli commands: they compose snapshot, press, and fill from the client process and own no daemon route, which is what keeps them out of the public catalog that requires one. Every nested call is an ordinary daemon command, so a policy-driven run carries the same claims, ref frames, and recording behaviour as a run an agent drives by hand, and each mutation is pinned to the generation of the snapshot that issued its ref. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both are pure delegators, so they join the `none` device-claim list. `act` joins the daemon-preserving and unbounded-envelope lists: a run is a sequence of ordinary commands that each carry their own envelope, and an outer one would abort a run still making progress while a reset would destroy the session the remaining steps need. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A command-reference section, a reference page covering the decision shape, the text-supply rule, the post-action checks, and the limitations, plus one paragraph on the skill routing card. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The changed-code quality gate flagged the step loop as critically complex and six exports as unconsumed. The loop now reads as snapshot, decide, resolve step, record, with the write and press branches and the keypad walk as their own functions, and the candidate projection resolves a node's role separately from building the candidate. The internal constants are module-local. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The hint read "Without it, Fall back to agent-driven policy" wherever it was interpolated. The clause now lives in its own module, lower-cased, so both the provider and the composer can name it without one importing the other. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A policy asked whether progress is blocked says yes on anything that looks like a wall. In live runs on an iOS sign-in flow the flag was set twice while the policy also named the element that led onward: the phone field on the sign-in screen, and "I'll start fresh" on an onboarding page offering a file import. The previous rule acted through the flag only above 0.9 confidence, so the second one ended the run three screens short of the goal. The flag is now terminal only together with nothing worth acting on: no target, or one below --min-confidence. A named target above the floor is acted on. Being wrong there costs one step that the same-screen check catches and the step budget bounds; believing the flag costs the run. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The CLI startup closure must not grow, and a static edge from the client facade pulled the whole policy subtree into every invocation. The policy head is optional, so its client loads inside the two methods that use it, the way debug symbols already does. Also replaces a `as string` on the chosen target with a discriminated judgment, and reaches the client fixture's device group through the spread idiom that file already uses for observability, so the oversized-test-file ratchet holds. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ady reaches The eager-closure gate holds src/cli.ts to the module count at the merge-base, and a command family is reached from it through the family registry, so a new family directory grows it by every module the family index evaluates. The gate names the remedy: give the new code a home in a module the closure already evaluates. suggest and act join the interaction family, whose vocabulary they already share: find, get, and is are reads that name an element, and act is a loop over the verbs beside it. Their CLI defaults move next to the flags that publish them, so the help prose and the loop read one declaration. The policy runtime stays under commands/policy and is still reached only through the client's on-demand import. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The move left both descriptors naming a deleted module, which the explain-command table caught, and left three exports with no consumers. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…nges Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The changed-line coverage gate passed at 83%, but the two largest gaps were the seam that pins a mutation to its snapshot generation and the text both commands print. A run that loses the pin fails silently, on whichever element inherited the ref, so it is worth its own test rather than coverage by a live run. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Owner
Author
|
Closing: this fork has workflows disabled pending the one-time manual enable, so no CI can run here. Verification is recorded on callstack#2654. |
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.
Fork-internal PR opened only to run the repository's CI workflows on this exact commit, because the upstream PR (callstack#2654) is gated on maintainer approval. Not for merge.