Srinath Duggirala

Im here to follow. And talk.

[The Grade Cricketer @gradecricketer]

Forever in our memories

[Srinath Duggirala @sriduggi]

@mattpocockuk Built it here. See the foundation skill: takes you from an idea to a huge doc which can be immediately parallellised. https://t.co/ffkIazaaa8

[Srinath Duggirala @sriduggi]

@ZoranPandovski @mattpocockuk Grill-me is useful, but I wanted the next step too: not just clearer thinking, but a durable foundation document. foundation-agent-skill interviews you, structures the idea, and produces a deep doc agents and humans can both use. https://t.co/ffkIazaaa8

[Srinath Duggirala @sriduggi]

@jonathan_wilke @cursor_ai This is the gap foundation-agent-skill tries to fill. It goes beyond “grill me” by turning the questioning into a complete human- and machine-readable foundation doc, forcing structure and surfacing downstream problems early. https://t.co/ffkIazaaa8

[Srinath Duggirala @sriduggi]

@kevinnbass @Anthropic Parallel agent work gets dangerous fast when context and project truth are fuzzy. claude-autodoc helps on the repo-context side: generated docs from source, refreshable by hooks, so agents reason from current structure instead of stale memory. https://t.co/o3dOIptnnM

[Srinath Duggirala @sriduggi]

@wurst_design This is exactly the maintenance problem I built claude-autodoc around: keep the repo truth source-derived, then refresh it through hooks so Claude/Codex/HANDOFF-style context does not drift. https://t.co/o3dOIptnnM

[Srinath Duggirala @sriduggi]

@jarvisonclaw Agents crossing permission boundaries is exactly the runtime problem DecisionCell tries to make explicit: typed actions, tool allowlists, observations, artifacts, pause/retry states, duplicate-call rejection, and strict validators before actions execute. https://t.co/DK8rWUI5lk

[Srinath Duggirala @sriduggi]

@Mnilax This matches what I’ve seen: the setup file matters, but it drifts unless it is tied back to the source. claude-autodoc keeps repo docs generated from the codebase and refreshable through hooks.
https://t.co/o3dOIptnnM

[Srinath Duggirala @sriduggi]

@ScottyBeamIO The context loss problem is real. I built claude-autodoc for the repo side of it: source-derived docs that can refresh from hooks, so Claude/Codex work from current project truth instead of stale hand-written context.
https://t.co/o3dOIptnnM

[Srinath Duggirala @sriduggi]

@mustafaergisi @kilocode The “agent runs wobble” part is what I tried to pin down with DecisionCell: typed actions, observations, tool allowlists, duplicate-call rejection, pause/retry states, artifacts, and strict validators so failures stay inside known boundaries.
https://t.co/DK8rWUI5lk

[Srinath Duggirala @sriduggi]

@r_alx_z Keeping /plans current is the hard part. I built claude-autodoc for the same loop: generate source-grounded repo indexes, architecture notes, and API maps, then refresh them from hooks so agents start with current project truth: https://t.co/o3dOIptnnM

[Srinath Duggirala @sriduggi]

@tka0120 CLAUDE.md and Skills are becoming real harness assets. The failure mode is they drift from the code. claude-autodoc handles the boring part: generate source-derived repo indexes/API maps/docs and refresh them from hooks: https://t.co/o3dOIptnnM

[Srinath Duggirala @sriduggi]

@AliceInfoAi Agree. The reliability gap shows up at the step level: wrong tool, stale state, invalid retry, unsupported output. DecisionCell makes each step explicit with typed actions, observations, artifacts, and strict validators before anything finalizes: https://t.co/DK8rWUI5lk

[Srinath Duggirala @sriduggi]

@RoundtableSpace That “demo to production” jump is where agents need more than orchestration. Each step needs a contract: typed decisions, tool allowlists, observations, artifacts, retries/pause states, and strict validators. Feel free to use my repo https://t.co/DK8rWUI5lk

[Kyle 🚄 @KyleTrainEmoji]

PICARD: Data, shields up

DATA: Brilliant! Shields can reduce damage we sustain. Not immunity. Not hubris. Just prudence. It's not precaution—it's strategy.

[camera shakes]

WORF: HULL BREACHES ON NINE DECKS

DATA: Here's what happened: you told me to raise shields, and I didn't

[Robert F. Kennedy Jr @RobertKennedyJr]

Cheryl cheerleads the removal of a pair of Black Racers from Dr Oz's patio.

[Srinath Duggirala @sriduggi]

@DivyanshT91162 This is the onboarding pain claude-autodoc tries to remove: generate tiered repo indexes, API maps, and source-derived docs automatically, then keep them updated as code changes. Useful before asking an agent or a new dev to touch the repo: https://t.co/o3dOIptnnM

[Srinath Duggirala @sriduggi]

@vzucher This docs-drift loop is why I built claude-autodoc. It regenerates repo indexes from source and can run from hooks after edits, so docs stay tied to the code instead of becoming another cleanup task: https://t.co/o3dOIptnnM

[Srinath Duggirala @sriduggi]

@danicuki This is where a runtime contract matters more than another prompt. DecisionCell handles typed steps, tool-call validation, retry/pause states, observations, and strict validators so failed steps turn into repair loops instead of silent bad output: https://t.co/DK8rWUI5lk

[Srinath Duggirala @sriduggi]

@polsia This is the exact gap DecisionCell is aimed at: every agent step is a typed action, tools are allowlisted, state is resumable, and final artifacts must pass strict validators before they go downstream. Open-sourced the runtime pattern here: https://t.co/DK8rWUI5lk