| name | naga-match |
| description | Generates a target artifact whose N4 cosine fidelity to the source fingerprint clears the per-(pattern-class, target-domain) p10 threshold from the N5 posterior. Loop: emit chunk, score, rewrite below threshold (cap 2 attempts), drop on third. Use when the user runs /naga:match or asks to generate a new artifact in the same shape as an existing one. Do not use for fingerprint-only reports (see /naga:fingerprint) or for cross-repo work (see /naga:match-across).
|
| model | sonnet |
| tools | ["Read","Write","Edit","Grep","Glob"] |
naga-match
Preconditions
- The source has been fingerprinted (
plugins/naga-observe/state/patterns/<hash>.json exists). If not, the skill auto-runs /naga:observe <source> first.
- The N5 posterior at
plugins/naga-learning/state/posterior.json is readable. Cold-start default: p10 = 0.6.
Inputs
- Slash command:
/naga:match <source> <target>
- Arguments:
source — path to the source artifact (must already be fingerprinted, or auto-fingerprinted here).target — path where the new artifact will be written.
Steps
- Resolve the source fingerprint. If missing, invoke
/naga:observe <source> and continue.
- Read the N5 posterior; pick the (pattern_class, target_domain) p10 threshold. Mark
cold_start: true when N=0.
- Spawn the naga-shaper Sonnet agent with
{source_fingerprint, threshold, target_path} and a structured-return clause.
- The shaper emits chunk-by-chunk; each chunk is scored against the source vector via N4 cosine; chunks below threshold get one rewrite, then are dropped.
- Aggregate chunk-level cosines into
(score, ci_low, ci_high, N) via bootstrap_ci.
- If source/target domains diverge (e.g.,
.py source -> .md target), the shaper escalates to naga-orchestrator Opus for the relaxation set; the shaper re-runs with the relaxed feature subset.
- Write the target artifact via
Write.
- Publish
naga.artifact.generated with {source_path, target_path, fidelity_score, ci_low, ci_high, N}.
Outputs
- The generated artifact at
<target>.
- One
naga.artifact.generated event on the bus.
- Console:
(score, ci_low, ci_high, N, chunks_dropped).
Handoff
If the developer wants an independent fidelity check, they run /naga:validate <target> <source> — the validator re-scores cold against the same fingerprint.
Failure modes
- F11 reward hacking — never weaken the N5 threshold to pass a chunk. The threshold is read-only here; only naga-learning writes it.
- F12 degeneration loop — same chunk rewrite loops below threshold. Hard cap: 2 attempts per chunk; drop on the third.
- F04 task drift — extending into prompt-engineering. Naga replicates; that's Wixie's surface.
