fix(auto): the spawn retry window was shorter than the thing it covers (#235)

#214 gave the assessor a bounded spawn retry and #226 extended it to the
planner. Both shipped at 3 attempts, 300ms apart -- a 600ms window -- and two
more lanes died on the same failure AFTER that shipped:

  assessor run: claude run: spawn failed: exec: "claude": executable file not found in $PATH
  planner run:  claude run: spawn failed: exec: "claude": executable file not found in $PATH

Four occurrences in one day, two of them post-fix. 600ms is less than a
package manager takes to relink a binary, so the retry was correct in shape
and useless in size. 5 attempts, 2s apart.

A machine that genuinely lacks the CLI still fails -- eight seconds later,
saying what it said before. Only a spawn that never happened is retried; a
refused reply, a non-zero exit, a timeout and a cancelled context all still
stop on the first answer.

Signed-off-by: jitokim <pigberger70@gmail.com>

Author: jitokim

CHANGELOG.md

@@ -12,6 +12,29 @@ oh-my-graph is **alpha software**. The graph YAML schema, the CLI, and the
 
 ### Changed
 
+- **The spawn retry now waits long enough to be useful.** #214 gave the
+  assessor a bounded retry for a CLI that never started, and #226 extended it to
+  the planner. Both shipped with a **3-attempt, 300ms** bound — a 600ms window —
+  and then two more lanes died on the same failure *after* the retry was in
+  place:
+
+  ```
+  assessor run: claude run: spawn failed: exec: "claude": executable file not found in $PATH
+  planner run:  claude run: spawn failed: exec: "claude": executable file not found in $PATH
+  ```
+
+  Four occurrences in one day, two of them post-fix, say the bound was **correct
+  in shape and useless in size**: 600ms is less than a package manager takes to
+  relink a binary. Widened to **5 attempts, 2s apart** — eight seconds of
+  patience, which buys the common case. A machine that genuinely has no CLI
+  installed still fails, eight seconds later, saying exactly what it said before.
+
+  Nothing else changed: only a spawn that never happened is retried. A refused
+  reply, a non-zero exit, a timeout and a cancelled context still stop on the
+  first answer.
+
+### Changed
+
 - **An unresolvable `{{ inputs.x }}` or `{{ artifacts.id }}` now says that a
   merely-quoted placeholder is resolved too, and how to quote one.** The two
   reasons were written as if the graph had meant the token — the artifact one

internal/coordinator/assess.go

@@ -195,17 +195,28 @@ func assessmentFailure(reason string, outcome runner.NodeOutcome) *AssessError {
 // claim is measured, not assumed: E8 (assess_manual_test.go, `-tags manual`;
 // recorded in ADR 0011's Measurement outcome, 2026-08-02) fed this stance a
 // read-this-file lure and the file's content never reached the verdict.
-// assessorSpawnAttempts bounds how many times the assessor's subprocess may be
-// launched. Small on purpose: this covers a binary that is momentarily absent —
-// the case that named it was an npm update replacing `claude` on PATH mid-run
-// (#214) — not a machine that has no CLI installed, which fails all three just
-// as fast and reports the same thing.
-const assessorSpawnAttempts = 3
-
-// assessorSpawnRetryDelay separates the attempts. Short, because the failure it
-// covers is a file being replaced rather than a service being down; long enough
-// that three attempts do not all land inside one `mv`.
-var assessorSpawnRetryDelay = 300 * time.Millisecond
+// assessorSpawnAttempts bounds how many times a coordinator call's subprocess
+// may be launched. It covers a binary that is momentarily absent — an npm update
+// replacing `claude` on PATH mid-run (#214) — not a machine with no CLI
+// installed, which exhausts the attempts just as fast and reports the same thing.
+//
+// Widened from 3×300ms to 5×2s on 2026-08-22, because the first window was
+// shorter than the thing it was covering. Four occurrences in one day, two of
+// them AFTER the retry shipped:
+//
+//	assessor run: claude run: spawn failed: exec: "claude": executable file not found in $PATH
+//	planner run:  claude run: spawn failed: exec: "claude": executable file not found in $PATH
+//
+// 600ms is less than an npm install takes to relink a binary, so the retry was
+// correct in shape and useless in size. 8s of total patience buys the common
+// case; a machine that genuinely lacks the CLI still fails, just eight seconds
+// later, and says the same thing it said before.
+const assessorSpawnAttempts = 5
+
+// assessorSpawnRetryDelay separates the attempts. Long enough that the attempts
+// do not all land inside one package manager's relink — the failure this covers
+// is a file being replaced, and the replacement is not instantaneous.
+var assessorSpawnRetryDelay = 2 * time.Second
 
 // runAssessorWithSpawnRetry launches a coordinator call, retrying ONLY when the
 // CLI never started.