Evidence-driven coding agent
Carthagent turns a software request into an acceptance contract, runs the checks that support it, and keeps the resulting evidence tied to the source that actually passed.
Carthagent is the next identity for the open-source project currently named pi2.
DELIVERY RECORD
RUN / 0248The problem
Generated code is not a delivery. A passing check is not current proof after the source it tested has changed.
Before implementation, Carthagent records the goal, artifacts, assumptions, acceptance criteria, and executable checks. The result is a contract that can be inspected instead of an unspoken definition of “done.”
invitation-expiry-testBoundary and elapsed-time behaviorCURRENT / PASSED
browser-invitation-flowReal interaction and browser-error captureCURRENT / PASSED
expired-link-recoveryPrevious result no longer matches current sourceSTALE / RERUN REQUIRED
Carthagent makes the delivery process inspectable. Wrong assumptions revise the plan. Failed checks enter a bounded repair loop. A handoff is verified only when required evidence is fresh.
Read the repository, installed APIs, constraints, and existing conventions.
Map acceptance criteria to concrete checks before editing product files.
Deliver a runnable vertical slice, then complete behavior and failure paths.
Run declared checks and bind their output to the current workspace fingerprint.
Optionally ask a detached, fresh-context reviewer to inspect the proposed change.
Report current evidence, launch instructions, review status, and limitations.
Failed or stale evidence returns to bounded repair and verification
Evidence is associated with the delivery run, contract revision, check definition, and workspace fingerprint. Relevant edits invalidate old results rather than leaving a green check attached to code it never exercised.
A result records which check ran, under which contract revision, against which source fingerprint.
When covered source changes, previously passing evidence becomes stale and must be rerun.
Failures can return to implementation, but automatic repair nudges and review attempts have explicit limits.
src/invitations/accept.ts +18 −7The acceptance handler changed after the previous browser check passed.
unit-invitationsCURRENTbrowser-invitation-flowSTALElintCURRENTRun a persistent coding session with visible plan, progress, checks, evidence, and model settings.
Use the same delivery contract in scripts or bounded non-interactive runs with explicit tool and time limits.
Review the active plan, acceptance mapping, check output, source fingerprint, and handoff state in a browser.
Every declared required check has fresh passing evidence for the current delivery contract and source fingerprint.
The checks are independently complete, every possible defect is excluded, or untrusted code is safely sandboxed. Human judgment still matters.
The next chapter of pi2
Carthagent is taking shape in public. The current implementation, evaluation history, and source remain available under the project’s existing name while the rename is prepared.
Explore pi2 on GitHub Return to top