v0.2.13

Maintenance release — repo moved to the nearform org; fixes package metadata to match npm provenance. No functional changes.

v0.2.11...v0.2.13

v0.2.12

Maintenance release — repo moved to the nearform org. No functional changes.

v0.2.11

What's changed

• GitHub tools: add an optional githubApiBaseUrl override (Octokit baseUrl), threaded through run → runner → loadGitHubExtension → buildGitHubTools → GitHubClient. Lets a real workflow run with its github_* calls pointed at a fake GitHub server for tests/evals. Falls back to GITHUB_API_URL; production leaves it unset (→ https://api.github.com). Adds GitHubClient baseUrl-injection tests.
• Docs: correct release trigger (published Release, not tag push).

Full Changelog: v0.2.10...v0.2.11

v0.2.10

Upgrade the underlying Pi SDK to 0.80.2 (from 0.79.8).

• Bump @earendil-works/pi-ai and @earendil-works/pi-coding-agent to ^0.80.2.
• Pi 0.80.0 moved the static model-catalog read off the pi-ai root; resolveModel() now uses getBuiltinModel from @earendil-works/pi-ai/providers/all instead of the deprecated root/compat getModel alias.
• No JSONL event-contract changes; full unit suite passes.

v0.2.9

Tunable maxSteps cap (#6)

Adds a step cap on agent runs — a "step" is one Pi turn (one model call plus the tool executions it triggers).

• Programmatic: run({ maxSteps }); RunResult.maxStepsReached reports whether the cap stopped the run.
• CLI: --max-steps <n>.
• When the agent completes maxSteps turns and still wants to continue, the run stops cleanly (agent_end, exit 0) and emits a gated max_steps_reached event. Suppressed when no cap is hit, so default-run JSONL stays byte-identical.

Pi exposes no max-turns / shouldStopAfterTurn seam through its SDK, so the cap is enforced from the event stream (count turn_end, abort when exceeded).

v0.2.8

Tunable rate-limit retry + bumped defaults

Pi auto-retries transient model errors (429 rate limit, 503 overloaded, 5xx, network/timeout) with exponential backoff, but its stock schedule (3 retries ≈ 14s) was too short for providers that limit on a per-minute window — notably Fireworks, whose TPM limits can take ~60s to clear, causing bursts to fail.

• Bumped default to 5 retries / 4s base → 4s, 8s, 16s, 32s, 64s (the final wait alone outlasts a 60s window).
• New flags --max-retries <n> (0 disables) and --retry-base-delay-ms <ms>; run() gains maxRetries / retryBaseDelayMs.
• Precedence: flag → operator ~/.pi/agent/settings.json retry block → agentic-pi default (operator config, incl. retry.provider.*, is preserved).
• Pauses remain observable via the existing auto_retry_start / auto_retry_end JSONL events.

See PR #5.

v0.2.7

Upgrade pi to 0.79.8 + Fireworks AI docs

• Bumped @earendil-works/pi-ai and @earendil-works/pi-coding-agent ^0.75.4 → ^0.79.8. The refreshed pi-ai catalog adds current Fireworks models (e.g. glm-5p2, deepseek-v4-pro, kimi-k2p7-code, minimax-m3) with accurate context/cost metadata. Type-compatible — no source changes required.
• Fireworks AI is a first-class model provider (Anthropic-compatible endpoint, FIREWORKS_API_KEY). Use --model fireworks/accounts/fireworks/models/<id>. README §5 documents the prefix, env var, curated model list, and the ~/.pi/agent/models.json escape hatch for uncurated ids.

See PR #4.

v0.2.6

Agent Skills support

agentic-pi now wires up the Agent Skills standard (the SKILL.md-in-a-folder convention from Claude Code), which Pi implements natively.

• --skill <path> (repeatable) + --no-skills — map an arbitrary skills folder (e.g. ~/.claude/skills) into the agent. --skill is additive even with --no-skills. Skills in Pi's default locations (.pi/skills/, .agents/skills/, ~/.pi/agent/skills/, ~/.agents/skills/, package skills/) are discovered automatically.
• Gated skills_status JSONL event — surfaces discovered skills (status, count, per-skill name/source/modelInvocable, mappedPaths). Suppressed on a default run with no skills, so existing JSONL fixtures stay byte-identical.
• run() gains skillPaths/noSkills options and a result.skills mirror for in-process callers.
• Docs: README section 11, AGENTS.md, CLAUDE.md.

See PR #3.

v0.2.5

Native OpenTelemetry traces + metrics export (`--otel`)

Opt-in OTEL integration that exports the Pi session as a span tree (`session → turn → {chat, execute_tool}`) plus metrics (tokens, cost, durations, tool/turn counts), using the standard OpenTelemetry JS SDK and standard `OTEL_*` env vars. Built for orchestrators (lastlight) that forward OTEL config into a sandboxed agentic-pi process.

• Off unless explicitly enabled (`--otel` or `AGENTIC_PI_OTEL_ENABLED=1`; `--no-otel` wins). A bare OTLP endpoint does not auto-enable.
• Metadata-only by default; raw content only via `--otel-include-content` (bounded + truncated).
• Safe by default: no-op + no SDK import when disabled; SDK diagnostics never reach stdout/stderr; an unreachable collector degrades to a warning, never fails the run.
• Reads W3C `TRACEPARENT` for cross-process trace correlation.

New flags: `--otel`, `--no-otel`, `--otel-include-content`, `--otel-service-name`, `--otel-endpoint`. See README section 10.

v0.2.4

What's Changed

• Bundle @ff-labs/pi-fff as default file search by @cliftonc in #2

Full Changelog: v0.2.3...v0.2.4