Configuration

pikopod reads pikopod.yaml from the working directory unless --config says otherwise. --config is accepted by every command. Run pikopod init to scaffold one. Unknown keys are startup errors, not silent no-ops. A typo’d or removed knob must never let you believe something is configured. If pikopod starts, every key you wrote is a key it understands.

A minimal working file

data_dir: ./pikopod-data
upstreams:
  examplepay:
    listen: /examplepay
    target: https://api.examplepay.com
    spec_source: https://api.examplepay.com/openapi.json
slack:
  webhook_url: https://hooks.slack.com/services/...

Environment variables override the file. The full set is in Environment variables.

Top-level keys

Key	Type	Default	Notes
`listen`	string	`127.0.0.1`	Bind address for both servers. Non-loopback requires a token.
`agent_port`	int	`4700`	The observing agent.
`sandbox_port`	int	`4600`	The sandbox.
`data_dir`	string	`pikopod-data`	Everything pikopod persists.
`token_file`	string		Listener token for non-loopback binds. Must not be group- or world-readable.
`tls`	object		`cert_file` and `key_file`. Both or neither.
`upstreams`	map		The one concept the whole tool turns on. See below.
`slack`	object		Alert delivery.
`llm`	object		Bring-your-own model key.
`warmup`	object	50 samples, 48 hours	Per-endpoint learning gates.
`refine`	object	off	Contract refinement from traffic.
`behaviour`	object	off	Observed state machine.
`sampling`	object	keep everything	Thin what recordings persist.
`retention`	object	size-only rotation	Age recordings and the event log out.
`spec_watch`	object	60 minutes	Re-check interval for declared drift.

upstreams

Each named upstream has a route on the agent and a target it forwards to. The target alone decides posture: a real provider means watch mode, a vendor sandbox means staging mode, the internal sandbox means scenario mode.

upstreams:
  examplepay:
    listen: /examplepay
    target: https://api.examplepay.com
    spec_source: https://api.examplepay.com/openapi.json
    volatile_fields: [request_id, timestamp]
    mute: ["/health"]
    incidents:
      client_errors: false
      client_error_rate: 0.05

Key	Default	Meaning
`listen`	`/<name>`	Route prefix on the agent. Must start with `/`, must not be `/` alone, and must not duplicate or nest inside another upstream’s route.
`target`	required	Absolute base URL this upstream forwards to, with scheme and host.
`volatile_fields`		Fields whose values are allowed to churn. See Volatile fields.
`mute`		Endpoint templates whose alerts are suppressed.
`spec_source`		Arms the declared-drift watcher: an `http(s)` URL, a local file, or `git:<ref>:<path>`. See Spec watch.
`incidents.client_errors`	`false`	Capture 4xx as incidents.
`incidents.client_error_rate`	`0.05`	Share of requests to one endpoint family that must be 4xx before one is reported. Below 20 requests no rate is claimed.

Upstream names may only contain letters, digits, _ and -, because the name becomes a file path component under data_dir.

listen and tokens

listen: 127.0.0.1
token_file: /etc/pikopod/token

Both servers bind loopback by default. Binding a non-loopback address requires a token. The agent sits in a production request path, and an unauthenticated listener on 0.0.0.0 is a way to lose data, so pikopod refuses to start rather than let it happen. Supply the token via PIKOPOD_TOKEN or token_file, never on the command line. The environment wins over the file. When a token is set, every request to either server must carry it in the X-Pikopod-Token header. Tokenless agent listeners additionally reject requests carrying a foreign Host header, which blocks DNS rebinding from a browser on the same machine.

tls

tls:
  cert_file: /etc/pikopod/cert.pem
  key_file: /etc/pikopod/key.pem

Serves both ports over HTTPS. Both files or neither. Self-signed certificates are fine; pikopod’s own CLI clients trust the configured certificate file directly and nothing else. To generate a pair:

openssl req -x509 -newkey ec -pkeyopt ec_paramgen_curve:prime256v1 \
  -keyout key.pem -out cert.pem -days 365 -nodes -subj /CN=pikopod

data_dir

data_dir: ./pikopod-data

Recordings, baselines, alert state, imported contracts, scenario packs, the event log and the sandbox store all live here. Back it up like state, not like a cache. Its layout is in Data directory.

sampling

sampling:
  rate: 0.25

rate thins what recordings persist, never what pikopod learns from. Every record still feeds the learner and the differ; the rate only gates the disk write. Error responses, drift-bearing records and pre-warmup traffic are always kept. 0 keeps guaranteed classes only, and is distinguishable from unset (1.0).

retention

retention:
  max_age_hours: 168

Ages recordings and the event log out of disk. A record is kept at least that long and deleted no later than roughly twice that age. 0 (default) keeps size-based rotation only. scenario from-drift and scenario reproduce read the event log, so fingerprints older than the window can no longer be pinned. Every incident row therefore carries reproducible until <time> and an export: hint. See Incident bundles.

slack

slack:
  webhook_url: https://hooks.slack.com/services/...
  min_level: WARN
  digest_hours: 24

Key	Meaning
`webhook_url`	A plain incoming webhook.
`min_level`	Delivery floor: `INFO` (default, deliver everything), `WARN`, or `ERR`. It floors the channel, not the record: muted alerts still appear in the local event log, in `pikopod status`, and in the digest.
`digest_hours`	Periodic digest of new findings by severity. `0` disables it.

Alert behaviour itself is a fixed contract, not a set of keys. See Alerts.

warmup

warmup:
  min_samples: 50
  min_hours: 48

pikopod watches an endpoint until it has seen min_samples responses across at least min_hours, then freezes a reference. Both are overridable for evaluation. Lowering them shortens the blind window and raises false positives. min_hours: 0 is explicitly distinguishable from unset. See Warmup and baselines.

spec_watch

spec_watch:
  interval_minutes: 60

Armed per upstream by spec_source. Fetches are ETag-gated and the pin never advances on its own; accepting a declared change is an explicit pikopod import <name> --update.

refine

refine:
  enabled: false
  prefer_spec: false

Off by default. When enabled, observed traffic accumulates an overlay beside the spec-derived contract, and matured observations join the effective contract. prefer_spec flips type-conflict precedence back to spec-wins; the default is traffic-wins once the sustain gates clear. Inspect the result with pikopod contract <sandbox>. See Contract.

behaviour

behaviour:
  enabled: false

Off by default. When enabled, the agent learns the provider’s observed state machine from traffic. See State machine.

llm

llm:
  provider: openai     # which model provider the key belongs to
  api_key: sk-...      # or set PIKOPOD_LLM_KEY in the environment
  model: gpt-4o-mini   # optional; each provider has a default

Bring your own key. pikopod never ships a key and never proxies your requests through anyone else. Set provider to the provider the key belongs to (openai or openrouter), and the key in api_key or in PIKOPOD_LLM_KEY. The provider’s own standard key variable is also honoured, so an existing environment keeps working. When a key is stored in pikopod.yaml, the file must be mode 0600. Environment-provided keys do not make the file secret. The key is optional. Exactly three things use it: plain-English scenario authoring (scenario create), pikopod fix, and importing from a documentation URL as the last resort after the deterministic rungs fail. Everything else works with no key at all.

quotas

Sandbox resource limits are fixed rather than configured:

Limit	Value
Stored resources per sandbox	10,000
Total storage	100 MiB
Single resource	256 KiB

Exceeding them returns 413 or 507 rather than degrading silently. Reset a sandbox’s state with pikopod sandbox reset <name>.

Sandboxes

Sandboxes are not a pikopod.yaml key. They are registered by pikopod import and stored in <data_dir>/sandboxes.json with their id, seed, spec source, origin, linked upstream, webhook URL and recordings-fallback flag. See Importing.

The init template

pikopod init writes this file with mode 0600 and refuses to overwrite an existing one:

# pikopod.yaml — see docs/config-reference.md for every key + default
# listen: 127.0.0.1        # non-loopback requires PIKOPOD_TOKEN (env) or token_file
# data_dir: pikopod-data

upstreams:
  # name each provider; your app's base URL points at http://127.0.0.1:4700/<name>
  # (any HTTP API works — the name and target below are placeholders)
  examplepay:
    target: https://api.examplepay.com
    # volatile_fields: [request_ref]   # extra per-provider normalization strips
    # mute: ["/transaction/verify/{id}"]

slack:
  # webhook_url: https://hooks.slack.com/services/…   # alerts land here

llm:
  # provider: openai   # the provider your key belongs to
  # api_key: sk-…      # for docs import, plain-English scenarios and pikopod fix; or PIKOPOD_LLM_KEY

# warmup:                    # eval-only overrides; defaults 50 samples / 48h
#   min_samples: 50
#   min_hours: 48

# refine:                    # contract refinement from observed traffic (off by default)
#   enabled: true            # traffic grows an overlay beside the spec contract
#   prefer_spec: false       # default: sustained traffic WINS type conflicts

# sampling:                  # thin what recordings PERSIST — learning always sees everything
#   rate: 0.1                # keep ~10% of routine records; errors/drift/pre-warmup always kept

# retention:                 # age recordings + the drift-event log out of disk
#   max_age_hours: 168       # kept >= this, deleted by ~2x this; 0 (default) = size-only rotation

Validation errors

Every configuration error names the key, the reason and the fix, and exits 2. The ones you are most likely to meet:

Error	Fix
`refusing to bind <addr> without a token`	Set `PIKOPOD_TOKEN` or `token_file`, or keep the default `127.0.0.1`.
`token file is readable by other users`	`chmod 600` the token file.
`pikopod.yaml contains an llm API key but is readable by other users`	`chmod 600 pikopod.yaml`, or move the key to the environment.
`tls needs both cert_file and key_file`	Set both, or neither.
`upstream listen routes overlap`	Give each upstream a distinct, non-nested route.
`upstream listen route cannot be /`	Use a named route like `/<name>`.
`sampling.rate must be between 0 and 1`	Use a fraction, or omit the key.
`configuration is not valid`	The message names the unknown key or the YAML syntax error.

pikopod doctor checks the file, the data directory, the salt, the ports and every upstream’s reachability in one pass. See Troubleshooting.

Getting started

Gate: catch declared changes in CI

Sandbox: rehearse failures

Scenarios

Observe: the agent

Reproduce and fix

Operations

Project

A minimal working file

Top-level keys

upstreams

listen and tokens

tls

data_dir

sampling

retention

slack

warmup

spec_watch

refine

behaviour

llm

quotas

Sandboxes

The init template

Validation errors

​A minimal working file

​Top-level keys

​upstreams

​listen and tokens

​tls

​data_dir

​sampling

​retention

​slack

​warmup

​spec_watch

​refine

​behaviour

​llm

​quotas

​Sandboxes

​The init template

​Validation errors

A minimal working file

Top-level keys

upstreams

listen and tokens

tls

data_dir

sampling

retention

slack

warmup

spec_watch

refine

behaviour

llm

quotas

Sandboxes

The init template

Validation errors