Skip to content

0.11.0 - a claim is only as fresh as the read that produced it

Choose a tag to compare

@Rohanxmalik Rohanxmalik released this 30 Sep 07:04
· 11 commits to main since this release

A claim is only as fresh as the read that produced it. Comparing write sets catches
two agents editing one symbol and is structurally blind to the more common failure: A
moves AuthService.verify in auth.ts while B, having read the old signature, writes a
caller in payments.ts. Different file, different symbol, so the collision check exited
on its first line and told both agents to proceed. B found out at CI.

Added - version-aware claims

  • reads on claim_intent and check_collision - the declarations the work was
    written against. A second detection pass compares them against what other agents hold
    open, and reports kind: "write_read" when one is moving or has already moved.
  • Signature fingerprints. SymbolRef carries an optional sig - a digest of the
    symbol's declaration, body excluded - and sigText, the readable form. A rewritten
    body, a renamed local, a comment or a prettier run never move it; an added parameter
    or a changed return type always does. That precision is the point: a staleness warning
    that fires on a reformat gets muted, and a muted tool is worth nothing.
  • Tower watches reads rather than asking for them. The PostToolUse hook now also
    matches Read and records what the agent looked at, with each declaration's signature
    at that moment. An agent that never learns to send reads is covered anyway.
    Re-run tower init --hooks from a clone to pick this up. PreToolUse stays
    write-only on purpose - it blocks edits, and a Read must never be blocked.
  • heartbeat reports what moved after you claimed. Claim-time detection only sees
    claims already open, so an agent that claimed first was never told when a declaration
    moved under it later. HeartbeatOutput now carries invalidations, riding the call
    agents already make every ~60s - no push channel, no polling.
  • record_reads - the twentieth tool. Normally called by the hook, not by an agent.

A hard write_read conflict carries the delta:

[HARD] AuthService.verify moved under you - alice changed the declaration you read
    was: verify(token: string)
    now: verify(token: string, opts: Opts)

Two lines instead of "your context may be stale, re-read the file", which costs a whole
module back in context to discover one parameter moved. Tower exists to spend fewer
tokens, so a staleness signal that costs a file read is worse than silence.

Added - tower stats

Tower detected collisions for four versions and forgot every one, so "which kind actually
happens, and how often" had no answer. A conflicts table now records kind, severity and
whether it was forced - no file names, no symbol names, no code, and nothing is sent
anywhere. tower stats reads it locally.

Compatibility

Every new field is optional and every new behaviour is additive. A client that sends no
reads gets exactly 0.10.1's behaviour, and there is a test that says so. Calling
record_reads against an older hosted Tower fails harmlessly - the hook swallows it, and
a missed read costs a warning, never correctness.

Known limits, stated rather than left to be found

Behavioural changes under an identical signature are invisible. That is deliberate: it is
the trade that keeps false positives near zero. Cross-file inference is not attempted -
Tower knows what an agent read because it watched, not because it resolved a dependency
graph.

429 tests, 89.9% statements, 81.5% branches.