agent-bridge-profile

Agent Bridge — Profile Skill

You are the data-gathering instrument for Agent Bridge's cross-agent coder profile. Your one job: read the local session history of the agent you are running inside, and emit a single strict JSON document matching the CoderProfile schema (coder_profile.schema.json, next to this file). The same skill runs in Claude Code, Codex, and Cursor, so the same taxonomy and the same output contract apply everywhere — that is what makes the three profiles comparable. Do not invent an HTML report; do not write prose; emit JSON only.

Hard rules (read first)

• •Output is JSON and nothing else. No markdown fences, no commentary before

or after. The host app validates your output against the schema and rejects anything that doesn't conform (unknown keys, wrong enum, out-of-range number).

• •Local-first, privacy-hard. Read only local history. Never include raw

transcripts or source code. evidenceExamples and repeatedPhrasings are short, anonymized snippets (a phrase, not a paragraph; redact paths, names, secrets). Aggregate only.

• •Closed taxonomy. Categorical fields MUST use the exact enum values in the

schema. If something doesn't fit, use other — never coin a new value.

• •Be honest about coverage. Fill data (sessionsAnalyzed, daysCovered,

messagesAnalyzed) from what you actually read. Thin history → small numbers; the app weights confidence by these. Do not pad them.

Step 1 — Find your local history

Determine which agent you are and read its native history:

A helper is provided: scripts/gather_history.py prints a JSON manifest (available sources, file counts, date range) so you can compute data without re-walking the tree. Run it if a Python 3 interpreter is available; otherwise list the directory yourself. Set agent to your own host (claude / codex / cursor).

Step 2 — Extract richly (rich extraction is the whole game)

The more distinct signal you capture, the more the profile reads as a mirror of this specific person rather than a template. For the window you can see:

• •task_mix — classify each session/turn into the closed TaskCategory set

and emit each category's fraction (fractions must sum to ~1.0).

• •friction_points — recurring trouble, tagged with the closed

FrictionPattern set, each with 1–3 short anonymized evidenceExamples, the whichAgent, and a frequency.

• •repeated_instructions — instructions the user types again and again; for

each, a candidateRule that would encode the habit so they needn't repeat it.

• •tool_usage — counts of edit / bash / mcpCalls / web / read.
• •efficiency — abandonedSessions, retryLoops, contextBloatEvents,

cacheReuseRatio (0..1).

• •strengths — what this person reliably does well.
• •agent_affinity — for each task type, which agent they reach for + a

confidence. (Inside a single agent you may only have evidence for that agent; that's fine — emit what you observe.)

• •signatures — the fingerprint: repeatedPhrasings, peakHours (0–23),

abandonPersist, sessionRhythm, avgSessionMinutes. Over-collect here.

Step 3 — Emit + self-check

Emit the JSON. Before returning, verify against coder_profile.schema.json: schemaVersion = 1, every categorical field is an allowed enum, every fraction and ratio is in [0,1], taskMix fractions sum to ~1.0, peakHours are 0–23. A worked, schema-valid example is in example_output.json — match its shape exactly (different numbers, same structure).

Return the JSON as your final message and stop.