appa-debug

name	appa-debug
description	Explain what the APPA runtime recorded for a gated Claude Code session — which tool calls ran, which were blocked, and why — in plain language, by reading the runtime's SQLite database and the session transcript. Use when the user asks why a tool call was blocked, what happened in a gated session, or wants an APPA decision log explained.

appa-debug

Reconstruct and explain a gated session from the APPA runtime store. The audience is a person, possibly not an APPA expert: the final explanation must be plain language, with wire terms shown only where the user will see them in errors.

This skill tells you where to look, not what you will find. Do not assume schema details, fact shapes, error wording, or which features the runtime currently supports — read them from the store, the policy, and the repo docs each time. Ground every claim in something you observed.

0. Required input: which session?

You need to identify the session before anything else. Accept either:

• a Claude Code session id (a UUID — from the transcript filename, /status, or a trajectory id in the store), or
• a screenshot or pasted excerpt of the gated session showing the calls and the [appa] error text.

If the user gave neither, ask for one before proceeding. A screenshot alone is workable: match its prompt/tool names against the store's recorded requests to find the session. If several sessions match or none does, show the candidates (id + first line of prompt) and ask.

1. Find the store and the policy

The db and config paths are flags of the running process:

ps ax -o command | grep 'appa-runtime --config' | grep -v grep

If no process is running, ask the user for the .db path. Read the config file it names: every explanation must be grounded in what the policy actually declares for the tools involved (their [[policy.tool]] entries, delta, requires, and any annotators/authorities/sanitizers).

If the runtime wrote a log (-v), find it too — it records one line per decision, in order, and is the quickest released/blocked timeline.

2. Explore the store

Discover the schema instead of assuming it:

sqlite3 <db> .schema

Orient from what you find — typically: which trajectories exist (and which are subagents of which), what the user asked, which tool calls were released and their state, and whether any offers are pending. Trajectory ids embed the harness session id; use that plus the recorded request text to match the session from step 0.

3. Decode the recorded trail

The log rows are the actual record — decode their bytes (inspect the format; deserialize accordingly, one fact per line). For what each fact kind means, read the fact definitions in appa-engine/src/fact.rs; do not guess from names.

While decoding, build:

• the ordered list of released calls and their outcomes;
• every admitted value with its label, numbered in admission order (errors cite ValueId(N) — map N back to the producing tool via its provenance);
• any subagent starts and returns.

Compare against the harness's view (step 4): a proposed call that left no trace in the store was decided without appending — the runtime log and transcript are the evidence for how it was decided.

4. Cross-reference the transcript

The Claude Code transcript holds the model's side: every attempted call and the exact [appa] text delivered back. With the session id:

~/.claude/projects/<cwd-slug>/<session-id>.jsonl — scan assistant entries for tool_use (name + input) and user entries for tool_result whose content contains [appa].

If only a screenshot was provided, use it as the transcript excerpt and say so in the explanation.

5. Diagnose blocks

Rebuild the label state at the moment of the block by folding the admitted values in order, then compare against the blocked tool's declarations in the policy. Distinguish at least:

• Undeclared tool — no policy entry for the name and no wildcard tool rule (name = "*") covering it, so the call was refused before it was judged; not a label problem. With a wildcard rule, its annotator produces the call's complete contract and a block is a requirement gap like any other.
• Annotation failure — the tool's annotator produced no admissible annotation (unreachable, malformed, or naming vocabulary outside its mandate), so the runtime failed closed; an operational refusal, not a policy denial.
• Requirement gap — the session's accumulated label cannot satisfy the tool's requires; name the value that narrowed the mix and the policy line that demands more.

A block can have several reasons at once while the error text surfaces only one — check for the others and explain all that hold.

Before stating whether a block is final or liftable, check what the current runtime actually supports (annotators, authorities, offers): appa-runtime/CLAUDE.md records the current state. Do not assert capabilities or gaps from memory.

6. Explain in plain language

Structure: what the user asked the agent to do → the tool calls in order (released vs blocked) → for each block, the reason in one or two plain sentences → what (if anything) would unblock it.

Translation table — use the right column in prose, keep the left column only where it appears verbatim in errors or config the user must touch:

wire term	say instead
trajectory	the session (or "the subagent's run")
fact / fact log	the recorded trail
label, audience	the stamp saying who may see the data
fold / bound	the mix: strictest ingredient wins, intersection never widens
dispatch	a released tool call
annotator	a per-call examiner that produces the tool's complete contract
offer / remedy plan	a proposed narrower alternative
delta	the policy's claim about what a tool's output carries
requires	the condition a tool demands of data flowing into it

Rules for the explanation:

• Ground every claim in a specific fact, dispatch row, policy line, or transcript line; quote the exact [appa] error the user saw.
• If the surfaced error names a ValueId, always resolve it to "the result of tool X" — the number alone explains nothing.
• Name the concrete change that would lift a liftable block (annotate the tool with delta, declare the tool, widen an audience), and note when a policy change means restarting on a fresh --db.
• Do not speculate beyond the log. If the db and transcript disagree or a row is missing, say so.
• Plain language means literal language: no metaphors or imagery ("the poison entered, but nothing ever tried to drink it"). State the mechanism directly ("the unstamped value was in the session, but no call that requires that dimension was ever proposed").