runtime: policy_path on the status read; a zero-rule baseline; the sync skill generates - #16
Draft
iskhakov wants to merge 12 commits into
Draft
runtime: policy_path on the status read; a zero-rule baseline; the sync skill generates#16iskhakov wants to merge 12 commits into
iskhakov wants to merge 12 commits into
Conversation
A second row shows tools: (every [[policy.tool]] entry) and rules:
(entries carrying more than the neutral bare-name-plus-delta={}
annotation), and names /appa-tool-sync as the skill that adds rules.
Counting is a line scan mirrored in statusline.ps1, verified against a
tomllib parse of the example policy, and fails open like the rest of
the statusline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The status wire answer gains policy_path, the canonical --config path: a fact of the process, not the log, so it joins the answer in the HTTP layer and never enters TrajectoryStatus. Both statuslines prefer that path for their tool counts and keep the APPA_CONFIG_DIR platform default as the fallback, so the statistics follow whichever runtime the session gates through -- the dev launch no longer needs APPA_CONFIG_DIR. docs/runtime.md's status-read contract carries the field. Also repairs plugin_hooks.rs, stale since the APPA_GATE guard: the drift check now pins SessionStart as the shared command with ensure-runtime.sh chained before the post, the harness runs gated, and a new test pins the ungated silent pass. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The examples name every built-in tool but no longer pre-fill
delta = {}: an unannotated tool's results are admitted as fully
unknown (UNK-5) until a reviewed rule gives the tool a real
annotation, which /appa-tool-sync generates. The neutral annotation
was a claim nobody had reviewed, made 40 times. The four real rules
stay: the Agent/Task run_in_background pins the runtime requires, and
the suspicious web results. The subagent crossing tests pin the
neutral annotation themselves -- they are about return crossings, not
label generation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The web tools lose their pre-filled suspicious deltas too: unknown already fails closed at every declared sink, and the suspicious rank is a reviewed decision the sync flow generates. The statusline's rules: counter stops counting parameters argument pins -- they constrain a call and label nothing -- so the Agent/Task pins the runtime requires no longer read as rules. A fresh deployment shows tools:44 rules:0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The interview collapses to three phases. Inventory walks the toolsets and MCP servers, skips a server that does not connect with one plan line instead of probing it, unpacks gateways, and never declares the runtime's reserved remedy tool. The plan is one plain-text message with one approval: ambiguity becomes a stated safer-side assumption the user flips in their reply -- no question dialogs, no pre-plan quiz. On approval, subagents generate in parallel: one per server drafting policy entries in the config's own dialect, one per dynamic resolver for tools whose sensitivity depends on an argument, learning the resolver contract from the machine. The generation reference the subagents need (the audience forms and the two loader rules) stays in the skill, out of the user's sight. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Phases 1 and 2 are mechanical: the lookups batch into two tool calls, the tool surface comes from the session context, and the skill forbids reading memory files or project docs -- recalled memory describes past machine states, every rule the run needs is written in the skill, and the machine is the source of truth. A minute of silence before the plan is named a failure. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The session context already says which servers answered and which need auth; claude mcp list health-probes everything and waits out the dead servers' timeouts, so it runs only when the context leaves a server's state genuinely unclear. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The shim gains --plugin-dir so gated sessions read every prompt file live from the checkout; the marketplace plugin is uninstalled meanwhile, since a plugin loaded twice fires every hook twice. 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>
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.
Cherry-picks the pending branch work onto this history (12 commits, conflicts resolved against what this main had already absorbed, texts translated to the protected vocabulary):
policy_pathon the status wire:GET /statusanswers carry the canonical--configpath, added in the HTTP layer and never enteringTrajectoryStatus; both statuslines prefer it for locating the policy they count, with theAPPA_CONFIG_DIR/platform default as fallback — the statistics follow whichever runtime the session is protected by, and the dev launch needs no extra variable.tools:/rules:counts, whererules:counts label rules only — an argument pin (parameters) constrains a call and labels nothing.delta— an unannotated tool's results are admitted as fully unknown until a reviewed rule says otherwise; the web tools' suspicious labels become generated decisions too. The subagent crossing tests pin the neutral annotation themselves (neutralized()), since they test return crossings, not label generation. A fresh deployment showstools:44 rules:0./appa-tool-syncbecomes inventory → one plan → generation: no question dialogs and no pre-plan quiz — ambiguity is a stated safer-side assumption the user flips in their reply; unreachable servers are skipped with one plan line; gateways unpack; the runtime's reserved remedy tool is never declared. On approval, subagents generate in parallel: policy entries per server, and a dynamic resolver per argument-dependent tool. The interview reads no memory and probes no server healths — the session context already knows.clappaat the checkout's plugin for live prompt files.Verified:
cargo test -p appa-runtime-v2green on this base (14 suites); this main's ownplugin_hooksrepair is kept where the histories disagreed.🤖 Generated with Claude Code